Upload 72 files

.gitattributes

@@ -44,3 +44,9 @@ hair_resposity/8.jpg filter=lfs diff=lfs merge=lfs -text
 test_imgs/ref2.jpg filter=lfs diff=lfs merge=lfs -text
 test_imgs/wjl.jpg filter=lfs diff=lfs merge=lfs -text
 test_imgs/zzf.jpg filter=lfs diff=lfs merge=lfs -text
+hair_resposity/6.jpg filter=lfs diff=lfs merge=lfs -text
+hair_resposity/7.jpg filter=lfs diff=lfs merge=lfs -text
+imgs/background.png filter=lfs diff=lfs merge=lfs -text
+imgs/multiview1.gif filter=lfs diff=lfs merge=lfs -text
+imgs/multiview2.gif filter=lfs diff=lfs merge=lfs -text
+imgs/teaser.jpg filter=lfs diff=lfs merge=lfs -text

ref_encoder/adapter.py

@@ -0,0 +1,40 @@
+import torch
+import torch.nn.functional as F
+
+def is_torch2_available():
+    return hasattr(F, "scaled_dot_product_attention")
+if is_torch2_available():
+    from .attention_processor import HairAttnProcessor2_0 as HairAttnProcessor, AttnProcessor2_0 as AttnProcessor
+else:
+    from .attention_processor import HairAttnProcessor, AttnProcessor
+
+def adapter_injection(unet, device="cuda", dtype=torch.float32, use_resampler=False, continue_learning_path=None):
+    device = device
+    dtype = dtype
+    # load Hair attention layers
+    attn_procs = {}
+    for name in unet.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+        if cross_attention_dim is None:
+            attn_procs[name] = HairAttnProcessor(hidden_size=hidden_size, cross_attention_dim=hidden_size, scale=1, use_resampler=use_resampler).to(device, dtype=dtype)
+        else:
+            attn_procs[name] = AttnProcessor()
+    unet.set_attn_processor(attn_procs)
+    adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
+    adapter_layers = adapter_modules
+    adapter_layers.to(device, dtype=dtype)
+
+    return adapter_layers
+
+def set_scale(unet, scale):
+    for attn_processor in unet.attn_processors.values():
+        if isinstance(attn_processor, HairAttnProcessor):
+            attn_processor.scale = scale

ref_encoder/attention.py

@@ -0,0 +1,481 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
+
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.models.attention import AdaLayerNorm, Attention, FeedForward
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from einops import rearrange
+from torch import nn
+
+
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(
+                dim, max_seq_length=num_positional_embeddings
+            )
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(
+                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+            )
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(
+                    dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+                )
+            )
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim
+                if not double_self_attention
+                else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(
+                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+            )
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(
+                dim, cross_attention_dim, num_attention_heads, attention_head_dim
+            )
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = (
+            cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        )
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states
+            if self.only_cross_attention
+            else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = (
+                norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+            )
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class TemporalBasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        name=None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
+        self.unet_use_temporal_attention = unet_use_temporal_attention
+        self.name=name
+
+        # SC-Attn
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+        self.norm1 = (
+            AdaLayerNorm(dim, num_embeds_ada_norm)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(dim)
+        )
+
+        # Cross-Attn
+        if cross_attention_dim is not None:
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            self.attn2 = None
+
+        if cross_attention_dim is not None:
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+        else:
+            self.norm2 = None
+
+        # Feed-forward
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.norm3 = nn.LayerNorm(dim)
+        self.use_ada_layer_norm_zero = False
+        # Temp-Attn
+        assert unet_use_temporal_attention is not None
+        if unet_use_temporal_attention:
+            self.attn_temp = Attention(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+            nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
+            self.norm_temp = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        timestep=None,
+        attention_mask=None,
+        video_length=None,
+        self_attention_additional_feats=None,
+        mode=None,
+    ):
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep)
+            if self.use_ada_layer_norm
+            else self.norm1(hidden_states)
+        )
+        if self.name:
+            modify_norm_hidden_states = norm_hidden_states
+            if mode == "write":
+                self_attention_additional_feats[self.name]=norm_hidden_states
+            elif mode == "read" and self_attention_additional_feats:
+                ref_states = self_attention_additional_feats[self.name]
+                bank_fea = [
+                    rearrange(
+                        ref_states.unsqueeze(1).repeat(1, video_length, 1, 1),
+                        "b t l c -> (b t) l c",
+                    )
+                ]
+                modify_norm_hidden_states = torch.cat(
+                    [norm_hidden_states] + bank_fea, dim=1
+                )
+
+            if self.unet_use_cross_frame_attention:
+                hidden_states = (
+                    self.attn1(
+                        norm_hidden_states,
+                        attention_mask=attention_mask,
+                        encoder_hidden_states=modify_norm_hidden_states,
+                        video_length=video_length,
+                    )
+                    + hidden_states
+                )
+            else:
+                hidden_states = (
+                    self.attn1(
+                        norm_hidden_states, 
+                        encoder_hidden_states=modify_norm_hidden_states,
+                        attention_mask=attention_mask
+                    )
+                    + hidden_states
+                )
+        else:    
+            if self.unet_use_cross_frame_attention:
+                hidden_states = (
+                    self.attn1(
+                        norm_hidden_states,
+                        attention_mask=attention_mask,
+                        video_length=video_length,
+                    )
+                    + hidden_states
+                )
+            else:
+                hidden_states = (
+                    self.attn1(norm_hidden_states, attention_mask=attention_mask)
+                    + hidden_states
+                )
+
+        if self.attn2 is not None:
+            # Cross-Attention
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm2(hidden_states)
+            )
+            hidden_states = (
+                self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                )
+                + hidden_states
+            )
+
+        # Feed-forward
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+        # Temporal-Attention
+        if self.unet_use_temporal_attention:
+            d = hidden_states.shape[1]
+            hidden_states = rearrange(
+                hidden_states, "(b f) d c -> (b d) f c", f=video_length
+            )
+            norm_hidden_states = (
+                self.norm_temp(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm_temp(hidden_states)
+            )
+            hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

ref_encoder/attention_processor.py

@@ -0,0 +1,391 @@
+# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.utils.import_utils import is_xformers_available
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+class HairAttnProcessor(nn.Module):
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, use_resampler=False):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.use_resampler = use_resampler
+        if self.use_resampler:
+            self.resampler = Resampler(query_dim=hidden_size)
+        self.to_k_SSR = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_SSR = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+            self,
+            attn,
+            hidden_states,
+            encoder_hidden_states=None,
+            attention_mask=None,
+            temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        # split hidden states
+        split_num = encoder_hidden_states.shape[1] // 2
+        encoder_hidden_states, _hidden_states = encoder_hidden_states[:, :split_num,
+                                                  :], encoder_hidden_states[:, split_num:, :]
+
+        if self.use_resampler:
+            _hidden_states = self.resampler(_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        _key = self.to_k_SSR(_hidden_states)
+        _value = self.to_v_SSR(_hidden_states)
+
+        _key = attn.head_to_batch_dim(_key)
+        _value = attn.head_to_batch_dim(_value)
+
+        _attention_probs = attn.get_attention_scores(query, _key, None)
+        _hidden_states = torch.bmm(_attention_probs, _value)
+        _hidden_states = attn.batch_to_head_dim(_hidden_states)
+
+        # # assume _hidden_states is a tensor of shape (batch_size, num_patches, hidden_size)
+        # batch_size, num_patches, hidden_size = _hidden_states.shape
+        # # create a mask tensor of shape (batch_size, num_patches)
+        # mask = torch.zeros((batch_size, num_patches), device="cuda", dtype=torch.float16)
+        # mask[:, 0:num_patches // 2] = 1
+        # # reshape the mask tensor to match the shape of _hidden_states
+        # mask = mask.unsqueeze(-1).expand(-1, -1, hidden_size)
+        # # apply the mask to _hidden_states
+        # _hidden_states = _hidden_states * mask
+
+        hidden_states = hidden_states + self.scale * _hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class HairAttnProcessor2_0(torch.nn.Module):
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, use_resampler=False):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.use_resampler = use_resampler
+        if self.use_resampler:
+            self.resampler = Resampler(query_dim=hidden_size)
+        self.to_k_SSR = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_SSR = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+            self,
+            attn,
+            hidden_states,
+            encoder_hidden_states=None,
+            attention_mask=None,
+            temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        # split hidden states
+        split_num = encoder_hidden_states.shape[1] // 2
+        encoder_hidden_states, _hidden_states = encoder_hidden_states[:, :split_num,
+                                                :], encoder_hidden_states[:, split_num:, :]
+
+        if self.use_resampler:
+            _hidden_states = self.resampler(_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        _hidden_states = _hidden_states.to(self.to_k_SSR.weight.dtype)
+        _key = self.to_k_SSR(_hidden_states)
+        _value = self.to_v_SSR(_hidden_states)
+
+        _key = _key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        _value = _value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        _hidden_states = F.scaled_dot_product_attention(
+            query.to(self.to_k_SSR.weight.dtype), _key, _value, attn_mask=None, dropout_p=0.0, is_causal=False
+        )
+
+        _hidden_states = _hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        _hidden_states = _hidden_states.to(query.dtype)
+
+        hidden_states = hidden_states + self.scale * _hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessor(nn.Module):
+    r"""
+    Default processor for performing attention-related computations.
+    """
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+class AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+            self,
+            hidden_size=None,
+            cross_attention_dim=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+            self,
+            attn,
+            hidden_states,
+            encoder_hidden_states=None,
+            attention_mask=None,
+            temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states

ref_encoder/motion_module.py

@@ -0,0 +1,388 @@
+# Adapt from https://github.com/guoyww/AnimateDiff/blob/main/animatediff/models/motion_module.py
+import math
+from dataclasses import dataclass
+from typing import Callable, Optional
+
+import torch
+from diffusers.models.attention import FeedForward
+from diffusers.models.attention_processor import Attention, AttnProcessor
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from einops import rearrange, repeat
+from torch import nn
+
+
+def zero_module(module):
+    # Zero out the parameters of a module and return it.
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+@dataclass
+class TemporalTransformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+def get_motion_module(in_channels, motion_module_type: str, motion_module_kwargs: dict):
+    if motion_module_type == "Vanilla":
+        return VanillaTemporalModule(
+            in_channels=in_channels,
+            **motion_module_kwargs,
+        )
+    else:
+        raise ValueError
+
+
+class VanillaTemporalModule(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads=8,
+        num_transformer_block=2,
+        attention_block_types=("Temporal_Self", "Temporal_Self"),
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+        temporal_attention_dim_div=1,
+        zero_initialize=True,
+    ):
+        super().__init__()
+
+        self.temporal_transformer = TemporalTransformer3DModel(
+            in_channels=in_channels,
+            num_attention_heads=num_attention_heads,
+            attention_head_dim=in_channels
+            // num_attention_heads
+            // temporal_attention_dim_div,
+            num_layers=num_transformer_block,
+            attention_block_types=attention_block_types,
+            cross_frame_attention_mode=cross_frame_attention_mode,
+            temporal_position_encoding=temporal_position_encoding,
+            temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+        )
+
+        if zero_initialize:
+            self.temporal_transformer.proj_out = zero_module(
+                self.temporal_transformer.proj_out
+            )
+
+    def forward(
+        self,
+        input_tensor,
+        temb,
+        encoder_hidden_states,
+        attention_mask=None,
+        anchor_frame_idx=None,
+    ):
+        hidden_states = input_tensor
+        hidden_states = self.temporal_transformer(
+            hidden_states, encoder_hidden_states, attention_mask
+        )
+
+        output = hidden_states
+        return output
+
+
+class TemporalTransformer3DModel(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads,
+        attention_head_dim,
+        num_layers,
+        attention_block_types=(
+            "Temporal_Self",
+            "Temporal_Self",
+        ),
+        dropout=0.0,
+        norm_num_groups=32,
+        cross_attention_dim=768,
+        activation_fn="geglu",
+        attention_bias=False,
+        upcast_attention=False,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+    ):
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
+        )
+        self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TemporalTransformerBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    attention_block_types=attention_block_types,
+                    dropout=dropout,
+                    norm_num_groups=norm_num_groups,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    attention_bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    cross_frame_attention_mode=cross_frame_attention_mode,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+                for d in range(num_layers)
+            ]
+        )
+        self.proj_out = nn.Linear(inner_dim, in_channels)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        assert (
+            hidden_states.dim() == 5
+        ), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        inner_dim = hidden_states.shape[1]
+        hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+            batch, height * weight, inner_dim
+        )
+        hidden_states = self.proj_in(hidden_states)
+
+        # Transformer Blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                video_length=video_length,
+            )
+
+        # output
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = (
+            hidden_states.reshape(batch, height, weight, inner_dim)
+            .permute(0, 3, 1, 2)
+            .contiguous()
+        )
+
+        output = hidden_states + residual
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+
+        return output
+
+
+class TemporalTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_attention_heads,
+        attention_head_dim,
+        attention_block_types=(
+            "Temporal_Self",
+            "Temporal_Self",
+        ),
+        dropout=0.0,
+        norm_num_groups=32,
+        cross_attention_dim=768,
+        activation_fn="geglu",
+        attention_bias=False,
+        upcast_attention=False,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+    ):
+        super().__init__()
+
+        attention_blocks = []
+        norms = []
+
+        for block_name in attention_block_types:
+            attention_blocks.append(
+                VersatileAttention(
+                    attention_mode=block_name.split("_")[0],
+                    cross_attention_dim=cross_attention_dim
+                    if block_name.endswith("_Cross")
+                    else None,
+                    query_dim=dim,
+                    heads=num_attention_heads,
+                    dim_head=attention_head_dim,
+                    dropout=dropout,
+                    bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    cross_frame_attention_mode=cross_frame_attention_mode,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+            )
+            norms.append(nn.LayerNorm(dim))
+
+        self.attention_blocks = nn.ModuleList(attention_blocks)
+        self.norms = nn.ModuleList(norms)
+
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.ff_norm = nn.LayerNorm(dim)
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        video_length=None,
+    ):
+        for attention_block, norm in zip(self.attention_blocks, self.norms):
+            norm_hidden_states = norm(hidden_states)
+            hidden_states = (
+                attention_block(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states
+                    if attention_block.is_cross_attention
+                    else None,
+                    video_length=video_length,
+                )
+                + hidden_states
+            )
+
+        hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states
+
+        output = hidden_states
+        return output
+
+
+class PositionalEncoding(nn.Module):
+    def __init__(self, d_model, dropout=0.0, max_len=24):
+        super().__init__()
+        self.dropout = nn.Dropout(p=dropout)
+        position = torch.arange(max_len).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model)
+        )
+        pe = torch.zeros(1, max_len, d_model)
+        pe[0, :, 0::2] = torch.sin(position * div_term)
+        pe[0, :, 1::2] = torch.cos(position * div_term)
+        self.register_buffer("pe", pe)
+
+    def forward(self, x):
+        x = x + self.pe[:, : x.size(1)]
+        return self.dropout(x)
+
+
+class VersatileAttention(Attention):
+    def __init__(
+        self,
+        attention_mode=None,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+        *args,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        assert attention_mode == "Temporal"
+
+        self.attention_mode = attention_mode
+        self.is_cross_attention = kwargs["cross_attention_dim"] is not None
+
+        self.pos_encoder = (
+            PositionalEncoding(
+                kwargs["query_dim"],
+                dropout=0.0,
+                max_len=temporal_position_encoding_max_len,
+            )
+            if (temporal_position_encoding and attention_mode == "Temporal")
+            else None
+        )
+
+    def extra_repr(self):
+        return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}"
+
+    def set_use_memory_efficient_attention_xformers(
+        self,
+        use_memory_efficient_attention_xformers: bool,
+        attention_op: Optional[Callable] = None,
+    ):
+        if use_memory_efficient_attention_xformers:
+            if not is_xformers_available():
+                raise ModuleNotFoundError(
+                    (
+                        "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+                        " xformers"
+                    ),
+                    name="xformers",
+                )
+            elif not torch.cuda.is_available():
+                raise ValueError(
+                    "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
+                    " only available for GPU "
+                )
+            else:
+                try:
+                    # Make sure we can run the memory efficient attention
+                    _ = xformers.ops.memory_efficient_attention(
+                        torch.randn((1, 2, 40), device="cuda"),
+                        torch.randn((1, 2, 40), device="cuda"),
+                        torch.randn((1, 2, 40), device="cuda"),
+                    )
+                except Exception as e:
+                    raise e
+
+            # XFormersAttnProcessor corrupts video generation and work with Pytorch 1.13.
+            # Pytorch 2.0.1 AttnProcessor works the same as XFormersAttnProcessor in Pytorch 1.13.
+            # You don't need XFormersAttnProcessor here.
+            # processor = XFormersAttnProcessor(
+            #     attention_op=attention_op,
+            # )
+            processor = AttnProcessor()
+        else:
+            processor = AttnProcessor()
+
+        self.set_processor(processor)
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        video_length=None,
+        **cross_attention_kwargs,
+    ):
+        if self.attention_mode == "Temporal":
+            d = hidden_states.shape[1]  # d means HxW
+            hidden_states = rearrange(
+                hidden_states, "(b f) d c -> (b d) f c", f=video_length
+            )
+
+            if self.pos_encoder is not None:
+                hidden_states = self.pos_encoder(hidden_states)
+
+            encoder_hidden_states = (
+                repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d)
+                if encoder_hidden_states is not None
+                else encoder_hidden_states
+            )
+
+        else:
+            raise NotImplementedError
+
+        hidden_states = self.processor(
+            self,
+            hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+        if self.attention_mode == "Temporal":
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

ref_encoder/mutual_self_attention.py

@@ -0,0 +1,365 @@
+# Adapted from https://github.com/magic-research/magic-animate/blob/main/magicanimate/models/mutual_self_attention.py
+from typing import Any, Dict, Optional
+
+import torch
+from einops import rearrange
+
+from src.models.attention import TemporalBasicTransformerBlock
+
+#from .attention import BasicTransformerBlock
+from diffusers.models.attention import BasicTransformerBlock
+
+def torch_dfs(model: torch.nn.Module):
+    result = [model]
+    for child in model.children():
+        result += torch_dfs(child)
+    return result
+
+
+class ReferenceAttentionControl:
+    def __init__(
+        self,
+        unet,
+        mode="write",
+        do_classifier_free_guidance=False,
+        attention_auto_machine_weight=float("inf"),
+        gn_auto_machine_weight=1.0,
+        style_fidelity=1.0,
+        reference_attn=True,
+        reference_adain=False,
+        fusion_blocks="midup",
+        batch_size=1,
+    ) -> None:
+        # 10. Modify self attention and group norm
+        self.unet = unet
+        assert mode in ["read", "write"]
+        assert fusion_blocks in ["midup", "full"]
+        self.reference_attn = reference_attn
+        self.reference_adain = reference_adain
+        self.fusion_blocks = fusion_blocks
+        self.register_reference_hooks(
+            mode,
+            do_classifier_free_guidance,
+            attention_auto_machine_weight,
+            gn_auto_machine_weight,
+            style_fidelity,
+            reference_attn,
+            reference_adain,
+            fusion_blocks,
+            batch_size=batch_size,
+        )
+
+    def register_reference_hooks(
+        self,
+        mode,
+        do_classifier_free_guidance,
+        attention_auto_machine_weight,
+        gn_auto_machine_weight,
+        style_fidelity,
+        reference_attn,
+        reference_adain,
+        dtype=torch.float16,
+        batch_size=1,
+        num_images_per_prompt=1,
+        device=torch.device("cpu"),
+        fusion_blocks="midup",
+    ):
+        MODE = mode
+        do_classifier_free_guidance = do_classifier_free_guidance
+        attention_auto_machine_weight = attention_auto_machine_weight
+        gn_auto_machine_weight = gn_auto_machine_weight
+        style_fidelity = style_fidelity
+        reference_attn = reference_attn
+        reference_adain = reference_adain
+        fusion_blocks = fusion_blocks
+        num_images_per_prompt = num_images_per_prompt
+        dtype = dtype
+        if do_classifier_free_guidance:
+            uc_mask = (
+                torch.Tensor(
+                    [1] * batch_size * num_images_per_prompt * 16
+                    + [0] * batch_size * num_images_per_prompt * 16
+                )
+                .to(device)
+                .bool()
+            )
+        else:
+            uc_mask = (
+                torch.Tensor([0] * batch_size * num_images_per_prompt * 2)
+                .to(device)
+                .bool()
+            )
+
+        def hacked_basic_transformer_inner_forward(
+            self,
+            hidden_states: torch.FloatTensor,
+            attention_mask: Optional[torch.FloatTensor] = None,
+            encoder_hidden_states: Optional[torch.FloatTensor] = None,
+            encoder_attention_mask: Optional[torch.FloatTensor] = None,
+            timestep: Optional[torch.LongTensor] = None,
+            cross_attention_kwargs: Dict[str, Any] = None,
+            class_labels: Optional[torch.LongTensor] = None,
+            video_length=None,
+            self_attention_additional_feats=None,
+            mode=None,
+        ):
+            if self.use_ada_layer_norm:  # False
+                norm_hidden_states = self.norm1(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero:
+                (
+                    norm_hidden_states,
+                    gate_msa,
+                    shift_mlp,
+                    scale_mlp,
+                    gate_mlp,
+                ) = self.norm1(
+                    hidden_states,
+                    timestep,
+                    class_labels,
+                    hidden_dtype=hidden_states.dtype,
+                )
+            else:
+                norm_hidden_states = self.norm1(hidden_states)
+
+            # 1. Self-Attention
+            # self.only_cross_attention = False
+            cross_attention_kwargs = (
+                cross_attention_kwargs if cross_attention_kwargs is not None else {}
+            )
+            if self.only_cross_attention:
+                attn_output = self.attn1(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states
+                    if self.only_cross_attention
+                    else None,
+                    attention_mask=attention_mask,
+                    **cross_attention_kwargs,
+                )
+            else:
+                if MODE == "write":
+                    self.bank.append(norm_hidden_states.clone())
+                    attn_output = self.attn1(
+                        norm_hidden_states,
+                        encoder_hidden_states=encoder_hidden_states
+                        if self.only_cross_attention
+                        else None,
+                        attention_mask=attention_mask,
+                        **cross_attention_kwargs,
+                    )
+                if MODE == "read":
+                    bank_fea = [
+                        rearrange(
+                            d.unsqueeze(1).repeat(1, video_length, 1, 1),
+                            "b t l c -> (b t) l c",
+                        )
+                        for d in self.bank
+                    ]
+                    modify_norm_hidden_states = torch.cat(
+                        [norm_hidden_states] + bank_fea, dim=1
+                    )
+                    hidden_states_uc = (
+                        self.attn1(
+                            norm_hidden_states,
+                            encoder_hidden_states=modify_norm_hidden_states,
+                            attention_mask=attention_mask,
+                        )
+                        + hidden_states
+                    )
+                    if do_classifier_free_guidance:
+                        hidden_states_c = hidden_states_uc.clone()
+                        _uc_mask = uc_mask.clone()
+                        if hidden_states.shape[0] != _uc_mask.shape[0]:
+                            _uc_mask = (
+                                torch.Tensor(
+                                    [1] * (hidden_states.shape[0] // 2)
+                                    + [0] * (hidden_states.shape[0] // 2)
+                                )
+                                .to(device)
+                                .bool()
+                            )
+                        hidden_states_c[_uc_mask] = (
+                            self.attn1(
+                                norm_hidden_states[_uc_mask],
+                                encoder_hidden_states=norm_hidden_states[_uc_mask],
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states[_uc_mask]
+                        )
+                        hidden_states = hidden_states_c.clone()
+                    else:
+                        hidden_states = hidden_states_uc
+
+                    # self.bank.clear()
+                    if self.attn2 is not None:
+                        # Cross-Attention
+                        norm_hidden_states = (
+                            self.norm2(hidden_states, timestep)
+                            if self.use_ada_layer_norm
+                            else self.norm2(hidden_states)
+                        )
+                        hidden_states = (
+                            self.attn2(
+                                norm_hidden_states,
+                                encoder_hidden_states=encoder_hidden_states,
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states
+                        )
+
+                    # Feed-forward
+                    hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+                    # Temporal-Attention
+                    if self.unet_use_temporal_attention:
+                        d = hidden_states.shape[1]
+                        hidden_states = rearrange(
+                            hidden_states, "(b f) d c -> (b d) f c", f=video_length
+                        )
+                        norm_hidden_states = (
+                            self.norm_temp(hidden_states, timestep)
+                            if self.use_ada_layer_norm
+                            else self.norm_temp(hidden_states)
+                        )
+                        hidden_states = (
+                            self.attn_temp(norm_hidden_states) + hidden_states
+                        )
+                        hidden_states = rearrange(
+                            hidden_states, "(b d) f c -> (b f) d c", d=d
+                        )
+
+                    return hidden_states
+
+            if self.use_ada_layer_norm_zero:
+                attn_output = gate_msa.unsqueeze(1) * attn_output
+            hidden_states = attn_output + hidden_states
+
+            if self.attn2 is not None:
+                norm_hidden_states = (
+                    self.norm2(hidden_states, timestep)
+                    if self.use_ada_layer_norm
+                    else self.norm2(hidden_states)
+                )
+
+                # 2. Cross-Attention
+                attn_output = self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=encoder_attention_mask,
+                    **cross_attention_kwargs,
+                )
+                hidden_states = attn_output + hidden_states
+
+            # 3. Feed-forward
+            norm_hidden_states = self.norm3(hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                norm_hidden_states = (
+                    norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+                )
+
+            ff_output = self.ff(norm_hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+            hidden_states = ff_output + hidden_states
+
+            return hidden_states
+
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            attn_modules = sorted(
+                attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+
+            for i, module in enumerate(attn_modules):
+                module._original_inner_forward = module.forward
+                if isinstance(module, BasicTransformerBlock):
+                    module.forward = hacked_basic_transformer_inner_forward.__get__(
+                        module, BasicTransformerBlock
+                    )
+                if isinstance(module, TemporalBasicTransformerBlock):
+                    module.forward = hacked_basic_transformer_inner_forward.__get__(
+                        module, TemporalBasicTransformerBlock
+                    )
+
+                module.bank = []
+                module.attn_weight = float(i) / float(len(attn_modules))
+
+    def update(self, writer, dtype=torch.float16):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, TemporalBasicTransformerBlock)
+                ]
+                writer_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(writer.unet.mid_block)
+                        + torch_dfs(writer.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, TemporalBasicTransformerBlock)
+                ]
+                writer_attn_modules = [
+                    module
+                    for module in torch_dfs(writer.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            reader_attn_modules = sorted(
+                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            writer_attn_modules = sorted(
+                writer_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            for r, w in zip(reader_attn_modules, writer_attn_modules):
+                r.bank = [v.clone().to(dtype) for v in w.bank]
+                # w.bank.clear()
+
+    def clear(self):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            reader_attn_modules = sorted(
+                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            for r in reader_attn_modules:
+                r.bank.clear()

ref_encoder/pose_guider.py

@@ -0,0 +1,57 @@
+from typing import Tuple
+
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.nn.init as init
+from diffusers.models.modeling_utils import ModelMixin
+
+from src.models.motion_module import zero_module
+from src.models.resnet import InflatedConv3d
+
+
+class PoseGuider(ModelMixin):
+    def __init__(
+        self,
+        conditioning_embedding_channels: int,
+        conditioning_channels: int = 3,
+        block_out_channels: Tuple[int] = (16, 32, 64, 128),
+    ):
+        super().__init__()
+        self.conv_in = InflatedConv3d(
+            conditioning_channels, block_out_channels[0], kernel_size=3, padding=1
+        )
+
+        self.blocks = nn.ModuleList([])
+
+        for i in range(len(block_out_channels) - 1):
+            channel_in = block_out_channels[i]
+            channel_out = block_out_channels[i + 1]
+            self.blocks.append(
+                InflatedConv3d(channel_in, channel_in, kernel_size=3, padding=1)
+            )
+            self.blocks.append(
+                InflatedConv3d(
+                    channel_in, channel_out, kernel_size=3, padding=1, stride=2
+                )
+            )
+
+        self.conv_out = zero_module(
+            InflatedConv3d(
+                block_out_channels[-1],
+                conditioning_embedding_channels,
+                kernel_size=3,
+                padding=1,
+            )
+        )
+
+    def forward(self, conditioning):
+        embedding = self.conv_in(conditioning)
+        embedding = F.silu(embedding)
+
+        for block in self.blocks:
+            embedding = block(embedding)
+            embedding = F.silu(embedding)
+
+        embedding = self.conv_out(embedding)
+
+        return embedding

ref_encoder/reference_control.py

@@ -0,0 +1,528 @@
+import torch
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+from diffusers.models.attention import BasicTransformerBlock
+from diffusers.models.unet_2d_blocks import CrossAttnDownBlock2D, CrossAttnUpBlock2D, DownBlock2D, UpBlock2D
+
+def torch_dfs(model: torch.nn.Module):
+    result = [model]
+    for child in model.children():
+        result += torch_dfs(child)
+    return result
+
+class ReferenceAttentionControl():
+
+    def __init__(self,
+                 unet,
+                 mode="write",
+                 do_classifier_free_guidance=False,
+                 attention_auto_machine_weight=float('inf'),
+                 gn_auto_machine_weight=1.0,
+                 style_fidelity=1.0,
+                 reference_attn=True,
+                 reference_adain=False,
+                 fusion_blocks="full",
+                 batch_size=1,
+                 ) -> None:
+        # 10. Modify self attention and group norm
+        self.unet = unet
+        assert mode in ["read", "write"]
+        assert fusion_blocks in ["midup", "full"]
+        self.reference_attn = reference_attn
+        self.reference_adain = reference_adain
+        self.fusion_blocks = fusion_blocks
+        self.register_reference_hooks(
+            mode,
+            do_classifier_free_guidance,
+            attention_auto_machine_weight,
+            gn_auto_machine_weight,
+            style_fidelity,
+            reference_attn,
+            reference_adain,
+            fusion_blocks,
+            batch_size=batch_size,
+        )
+
+    def register_reference_hooks(
+            self,
+            mode,
+            do_classifier_free_guidance,
+            attention_auto_machine_weight,
+            gn_auto_machine_weight,
+            style_fidelity,
+            reference_attn,
+            reference_adain,
+            dtype=torch.float16,
+            batch_size=1,
+            num_images_per_prompt=1,
+            device=torch.device("cpu"),
+            fusion_blocks='midup',
+    ):
+        MODE = mode
+        do_classifier_free_guidance = do_classifier_free_guidance
+        attention_auto_machine_weight = attention_auto_machine_weight
+        gn_auto_machine_weight = gn_auto_machine_weight
+        style_fidelity = style_fidelity
+        reference_attn = reference_attn
+        reference_adain = reference_adain
+        fusion_blocks = fusion_blocks
+        num_images_per_prompt = num_images_per_prompt
+        dtype = dtype
+        if do_classifier_free_guidance:
+            uc_mask = (
+                torch.Tensor(
+                    [1] * batch_size * num_images_per_prompt * 16 + [0] * batch_size * num_images_per_prompt * 16)
+                .to(device)
+                .bool()
+            )
+        else:
+            uc_mask = (
+                torch.Tensor([0] * batch_size * num_images_per_prompt * 2)
+                .to(device)
+                .bool()
+            )
+
+        def hacked_basic_transformer_inner_forward(
+                self,
+                hidden_states: torch.FloatTensor,
+                attention_mask: Optional[torch.FloatTensor] = None,
+                encoder_hidden_states: Optional[torch.FloatTensor] = None,
+                encoder_attention_mask: Optional[torch.FloatTensor] = None,
+                timestep: Optional[torch.LongTensor] = None,
+                cross_attention_kwargs: Dict[str, Any] = None,
+                class_labels: Optional[torch.LongTensor] = None,
+        ):
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm1(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero:
+                norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                    hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+                )
+            else:
+                norm_hidden_states = self.norm1(hidden_states)
+
+            # 1. Self-Attention
+            cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+            if self.only_cross_attention:
+                attn_output = self.attn1(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+                    attention_mask=attention_mask,
+                    **cross_attention_kwargs,
+                )
+            else:
+                if MODE == "write":
+                    self.bank.append(norm_hidden_states.clone())
+                    attn_output = self.attn1(
+                        norm_hidden_states,
+                        encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+                        attention_mask=attention_mask,
+                        **cross_attention_kwargs,
+                    )
+                if MODE == "read":
+                    hidden_states_uc = self.attn1(norm_hidden_states,
+                                                  encoder_hidden_states=torch.cat([norm_hidden_states] + self.bank,
+                                                                                  dim=1),
+                                                  attention_mask=attention_mask) + hidden_states
+                    hidden_states_c = hidden_states_uc.clone()
+                    _uc_mask = uc_mask.clone()
+                    if do_classifier_free_guidance:
+                        if hidden_states.shape[0] != _uc_mask.shape[0]:
+                            _uc_mask = (
+                                torch.Tensor([1] * (hidden_states.shape[0] // 2) + [0] * (hidden_states.shape[0] // 2))
+                                .to(device)
+                                .bool()
+                            )
+
+                        hidden_states_c[_uc_mask] = self.attn1(
+                            norm_hidden_states[_uc_mask],
+                            encoder_hidden_states=norm_hidden_states[_uc_mask],
+                            attention_mask=attention_mask,
+                        ) + hidden_states[_uc_mask]
+                    hidden_states = hidden_states_c.clone()
+
+                    self.bank.clear()
+                    if self.attn2 is not None:
+                        # Cross-Attention
+                        norm_hidden_states = (
+                            self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(
+                                hidden_states)
+                        )
+                        hidden_states = (
+                                self.attn2(
+                                    norm_hidden_states, encoder_hidden_states=encoder_hidden_states,
+                                    attention_mask=attention_mask
+                                )
+                                + hidden_states
+                        )
+
+                    # Feed-forward
+                    hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+                    return hidden_states
+
+            if self.use_ada_layer_norm_zero:
+                attn_output = gate_msa.unsqueeze(1) * attn_output
+            hidden_states = attn_output + hidden_states
+
+            if self.attn2 is not None:
+                norm_hidden_states = (
+                    self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+                )
+
+                # 2. Cross-Attention
+                attn_output = self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=encoder_attention_mask,
+                    **cross_attention_kwargs,
+                )
+                hidden_states = attn_output + hidden_states
+
+            # 3. Feed-forward
+            norm_hidden_states = self.norm3(hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+            ff_output = self.ff(norm_hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+            hidden_states = ff_output + hidden_states
+
+            return hidden_states
+
+        def hacked_mid_forward(self, *args, **kwargs):
+            eps = 1e-6
+            x = self.original_forward(*args, **kwargs)
+            if MODE == "write":
+                if gn_auto_machine_weight >= self.gn_weight:
+                    var, mean = torch.var_mean(x, dim=(2, 3), keepdim=True, correction=0)
+                    self.mean_bank.append(mean)
+                    self.var_bank.append(var)
+            if MODE == "read":
+                if len(self.mean_bank) > 0 and len(self.var_bank) > 0:
+                    var, mean = torch.var_mean(x, dim=(2, 3), keepdim=True, correction=0)
+                    std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
+                    mean_acc = sum(self.mean_bank) / float(len(self.mean_bank))
+                    var_acc = sum(self.var_bank) / float(len(self.var_bank))
+                    std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
+                    x_uc = (((x - mean) / std) * std_acc) + mean_acc
+                    x_c = x_uc.clone()
+                    if do_classifier_free_guidance and style_fidelity > 0:
+                        x_c[uc_mask] = x[uc_mask]
+                    x = style_fidelity * x_c + (1.0 - style_fidelity) * x_uc
+                self.mean_bank = []
+                self.var_bank = []
+            return x
+
+        def hack_CrossAttnDownBlock2D_forward(
+                self,
+                hidden_states: torch.FloatTensor,
+                temb: Optional[torch.FloatTensor] = None,
+                encoder_hidden_states: Optional[torch.FloatTensor] = None,
+                attention_mask: Optional[torch.FloatTensor] = None,
+                cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+                encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        ):
+            eps = 1e-6
+
+            # TODO(Patrick, William) - attention face_hair_mask is not used
+            output_states = ()
+
+            for i, (resnet, attn) in enumerate(zip(self.resnets, self.attentions)):
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )[0]
+                if MODE == "write":
+                    if gn_auto_machine_weight >= self.gn_weight:
+                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
+                        self.mean_bank.append([mean])
+                        self.var_bank.append([var])
+                if MODE == "read":
+                    if len(self.mean_bank) > 0 and len(self.var_bank) > 0:
+                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
+                        std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
+                        mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))
+                        var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))
+                        std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
+                        hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc
+                        hidden_states_c = hidden_states_uc.clone()
+                        if do_classifier_free_guidance and style_fidelity > 0:
+                            hidden_states_c[uc_mask] = hidden_states[uc_mask].to(hidden_states_c.dtype)
+                        hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc
+
+                output_states = output_states + (hidden_states,)
+
+            if MODE == "read":
+                self.mean_bank = []
+                self.var_bank = []
+
+            if self.downsamplers is not None:
+                for downsampler in self.downsamplers:
+                    hidden_states = downsampler(hidden_states)
+
+                output_states = output_states + (hidden_states,)
+
+            return hidden_states, output_states
+
+        def hacked_DownBlock2D_forward(self, hidden_states, temb=None):
+            eps = 1e-6
+
+            output_states = ()
+
+            for i, resnet in enumerate(self.resnets):
+                hidden_states = resnet(hidden_states, temb)
+
+                if MODE == "write":
+                    if gn_auto_machine_weight >= self.gn_weight:
+                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
+                        self.mean_bank.append([mean])
+                        self.var_bank.append([var])
+                if MODE == "read":
+                    if len(self.mean_bank) > 0 and len(self.var_bank) > 0:
+                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
+                        std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
+                        mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))
+                        var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))
+                        std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
+                        hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc
+                        hidden_states_c = hidden_states_uc.clone()
+                        if do_classifier_free_guidance and style_fidelity > 0:
+                            hidden_states_c[uc_mask] = hidden_states[uc_mask].to(hidden_states_c.dtype)
+                        hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc
+
+                output_states = output_states + (hidden_states,)
+
+            if MODE == "read":
+                self.mean_bank = []
+                self.var_bank = []
+
+            if self.downsamplers is not None:
+                for downsampler in self.downsamplers:
+                    hidden_states = downsampler(hidden_states)
+
+                output_states = output_states + (hidden_states,)
+
+            return hidden_states, output_states
+
+        def hacked_CrossAttnUpBlock2D_forward(
+                self,
+                hidden_states: torch.FloatTensor,
+                res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+                temb: Optional[torch.FloatTensor] = None,
+                encoder_hidden_states: Optional[torch.FloatTensor] = None,
+                cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+                upsample_size: Optional[int] = None,
+                attention_mask: Optional[torch.FloatTensor] = None,
+                encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        ):
+            eps = 1e-6
+            # TODO(Patrick, William) - attention face_hair_mask is not used
+            for i, (resnet, attn) in enumerate(zip(self.resnets, self.attentions)):
+                # pop res hidden states
+                res_hidden_states = res_hidden_states_tuple[-1]
+                res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+                hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )[0]
+
+                if MODE == "write":
+                    if gn_auto_machine_weight >= self.gn_weight:
+                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
+                        self.mean_bank.append([mean])
+                        self.var_bank.append([var])
+                if MODE == "read":
+                    if len(self.mean_bank) > 0 and len(self.var_bank) > 0:
+                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
+                        std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
+                        mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))
+                        var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))
+                        std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
+                        hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc
+                        hidden_states_c = hidden_states_uc.clone()
+                        if do_classifier_free_guidance and style_fidelity > 0:
+                            hidden_states_c[uc_mask] = hidden_states[uc_mask].to(hidden_states_c.dtype)
+                        hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc
+
+            if MODE == "read":
+                self.mean_bank = []
+                self.var_bank = []
+
+            if self.upsamplers is not None:
+                for upsampler in self.upsamplers:
+                    hidden_states = upsampler(hidden_states, upsample_size)
+
+            return hidden_states
+
+        def hacked_UpBlock2D_forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
+            eps = 1e-6
+            for i, resnet in enumerate(self.resnets):
+                # pop res hidden states
+                res_hidden_states = res_hidden_states_tuple[-1]
+                res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+                hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+                hidden_states = resnet(hidden_states, temb)
+
+                if MODE == "write":
+                    if gn_auto_machine_weight >= self.gn_weight:
+                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
+                        self.mean_bank.append([mean])
+                        self.var_bank.append([var])
+                if MODE == "read":
+                    if len(self.mean_bank) > 0 and len(self.var_bank) > 0:
+                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
+                        std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
+                        mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))
+                        var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))
+                        std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
+                        hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc
+                        hidden_states_c = hidden_states_uc.clone()
+                        if do_classifier_free_guidance and style_fidelity > 0:
+                            hidden_states_c[uc_mask] = hidden_states[uc_mask].to(hidden_states_c.dtype)
+                        hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc
+
+            if MODE == "read":
+                self.mean_bank = []
+                self.var_bank = []
+
+            if self.upsamplers is not None:
+                for upsampler in self.upsamplers:
+                    hidden_states = upsampler(hidden_states, upsample_size)
+
+            return hidden_states
+
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                attn_modules = [module for module in (torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks))
+                                if isinstance(module, BasicTransformerBlock)]
+            elif self.fusion_blocks == "full":
+                attn_modules = [module for module in torch_dfs(self.unet) if isinstance(module, BasicTransformerBlock)]
+            attn_modules = sorted(attn_modules, key=lambda x: -x.norm1.normalized_shape[0])
+
+            for i, module in enumerate(attn_modules):
+                module._original_inner_forward = module.forward
+                module.forward = hacked_basic_transformer_inner_forward.__get__(module, BasicTransformerBlock)
+                module.bank = []
+                module.attn_weight = float(i) / float(len(attn_modules))
+
+        if self.reference_adain:
+            gn_modules = [self.unet.mid_block]
+            self.unet.mid_block.gn_weight = 0
+
+            down_blocks = self.unet.down_blocks
+            for w, module in enumerate(down_blocks):
+                module.gn_weight = 1.0 - float(w) / float(len(down_blocks))
+                gn_modules.append(module)
+
+            up_blocks = self.unet.up_blocks
+            for w, module in enumerate(up_blocks):
+                module.gn_weight = float(w) / float(len(up_blocks))
+                gn_modules.append(module)
+
+            for i, module in enumerate(gn_modules):
+                if getattr(module, "original_forward", None) is None:
+                    module.original_forward = module.forward
+                if i == 0:
+                    # mid_block
+                    module.forward = hacked_mid_forward.__get__(module, torch.nn.Module)
+                elif isinstance(module, CrossAttnDownBlock2D):
+                    module.forward = hack_CrossAttnDownBlock2D_forward.__get__(module, CrossAttnDownBlock2D)
+                elif isinstance(module, DownBlock2D):
+                    module.forward = hacked_DownBlock2D_forward.__get__(module, DownBlock2D)
+                elif isinstance(module, CrossAttnUpBlock2D):
+                    module.forward = hacked_CrossAttnUpBlock2D_forward.__get__(module, CrossAttnUpBlock2D)
+                elif isinstance(module, UpBlock2D):
+                    module.forward = hacked_UpBlock2D_forward.__get__(module, UpBlock2D)
+                module.mean_bank = []
+                module.var_bank = []
+                module.gn_weight *= 2
+
+    def update(self, writer, dtype=torch.float16):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [module for module in
+                                       (torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)) if
+                                       isinstance(module, BasicTransformerBlock)]
+                writer_attn_modules = [module for module in
+                                       (torch_dfs(writer.unet.mid_block) + torch_dfs(writer.unet.up_blocks)) if
+                                       isinstance(module, BasicTransformerBlock)]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [module for module in torch_dfs(self.unet) if
+                                       isinstance(module, BasicTransformerBlock)]
+                writer_attn_modules = [module for module in torch_dfs(writer.unet) if
+                                       isinstance(module, BasicTransformerBlock)]
+            reader_attn_modules = sorted(reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0])
+            writer_attn_modules = sorted(writer_attn_modules, key=lambda x: -x.norm1.normalized_shape[0])
+            for r, w in zip(reader_attn_modules, writer_attn_modules):
+                r.bank = [v.clone().to(dtype) for v in w.bank]
+
+        if self.reference_adain:
+            reader_gn_modules = [self.unet.mid_block]
+
+            down_blocks = self.unet.down_blocks
+            for w, module in enumerate(down_blocks):
+                reader_gn_modules.append(module)
+
+            up_blocks = self.unet.up_blocks
+            for w, module in enumerate(up_blocks):
+                reader_gn_modules.append(module)
+
+            writer_gn_modules = [writer.unet.mid_block]
+
+            down_blocks = writer.unet.down_blocks
+            for w, module in enumerate(down_blocks):
+                writer_gn_modules.append(module)
+
+            up_blocks = writer.unet.up_blocks
+            for w, module in enumerate(up_blocks):
+                writer_gn_modules.append(module)
+
+            for r, w in zip(reader_gn_modules, writer_gn_modules):
+                if len(w.mean_bank) > 0 and isinstance(w.mean_bank[0], list):
+                    r.mean_bank = [[v.clone().to(dtype) for v in vl] for vl in w.mean_bank]
+                    r.var_bank = [[v.clone().to(dtype) for v in vl] for vl in w.var_bank]
+                else:
+                    r.mean_bank = [v.clone().to(dtype) for v in w.mean_bank]
+                    r.var_bank = [v.clone().to(dtype) for v in w.var_bank]
+
+    def clear(self):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [module for module in
+                                       (torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)) if
+                                       isinstance(module, BasicTransformerBlock)]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [module for module in torch_dfs(self.unet) if
+                                       isinstance(module, BasicTransformerBlock)]
+            reader_attn_modules = sorted(reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0])
+            for r in reader_attn_modules:
+                r.bank.clear()
+        if self.reference_adain:
+            reader_gn_modules = [self.unet.mid_block]
+
+            down_blocks = self.unet.down_blocks
+            for w, module in enumerate(down_blocks):
+                reader_gn_modules.append(module)
+
+            up_blocks = self.unet.up_blocks
+            for w, module in enumerate(up_blocks):
+                reader_gn_modules.append(module)
+
+            for r in reader_gn_modules:
+                r.mean_bank.clear()
+                r.var_bank.clear()

ref_encoder/reference_unet.py

@@ -0,0 +1,1053 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.lora import LoRALinearLayer
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    PositionNet,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.unet_2d_blocks import (
+    UNetMidBlock2DCrossAttn,
+    UNetMidBlock2DSimpleCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class CCProjection(ModelMixin, ConfigMixin):
+    def __init__(self, in_channel=772, out_channel=768):
+        super().__init__()
+        self.in_channel = in_channel
+        self.out_channel = out_channel
+        self.projection = torch.nn.Linear(in_channel, out_channel)
+
+    def forward(self, x):
+        return self.projection(x)
+
+
+class Identity(torch.nn.Module):
+    def __init__(self, scale=None, *args, **kwargs) -> None:
+        super(Identity, self).__init__()
+    def forward(self, input, *args, **kwargs):
+        return input
+
+
+class _LoRACompatibleLinear(nn.Module):
+    """
+    A Linear layer that can be used with LoRA.
+    """
+
+    def __init__(self, *args, lora_layer: Optional[LoRALinearLayer] = None, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.lora_layer = lora_layer
+
+    def set_lora_layer(self, lora_layer: Optional[LoRALinearLayer]):
+        self.lora_layer = lora_layer
+
+    def _fuse_lora(self):
+        pass
+
+    def _unfuse_lora(self):
+        pass
+
+    def forward(self, hidden_states, scale=None, lora_scale: int = 1):
+        return hidden_states
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+
+
+class ref_unet(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be either `UNetMidBlock2DCrossAttn` or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
+        conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
+        projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: int = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads=64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+        
+        self.add_time_proj = Timesteps(256, True, downscale_freq_shift=0)
+        self.add_angle_proj = Timesteps(512, True, downscale_freq_shift=0)  # encode camera angles
+        self.add_embedding = TimestepEmbedding(1280, 1280)
+
+        # time
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                attention_type=attention_type,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                cross_attention_dim=cross_attention_dim[-1],
+                attention_head_dim=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                skip_time_act=resnet_skip_time_act,
+                only_cross_attention=mid_block_only_cross_attention,
+                cross_attention_norm=cross_attention_norm,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+        self.up_blocks[3].attentions[2].transformer_blocks[0].attn1.to_q = _LoRACompatibleLinear()
+        self.up_blocks[3].attentions[2].transformer_blocks[0].attn1.to_k = _LoRACompatibleLinear()
+        self.up_blocks[3].attentions[2].transformer_blocks[0].attn1.to_v = _LoRACompatibleLinear()
+        self.up_blocks[3].attentions[2].transformer_blocks[0].attn1.to_out = nn.ModuleList([Identity(), Identity()])
+        self.up_blocks[3].attentions[2].transformer_blocks[0].norm2 = Identity()
+        self.up_blocks[3].attentions[2].transformer_blocks[0].attn2 = None
+        self.up_blocks[3].attentions[2].transformer_blocks[0].norm3 = Identity()
+        self.up_blocks[3].attentions[2].transformer_blocks[0].ff = Identity()
+        self.up_blocks[3].attentions[2].proj_out = Identity()
+
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = PositionNet(
+                positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        add_time_ids: List=None,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention face_hair_mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the face_hair_mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb, hint = self.add_embedding(image_embs, hint)
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+        
+        if add_time_ids is not None:
+            cond_aug, polars, azimuths = add_time_ids
+            cond_aug_emb = self.add_time_proj(cond_aug.flatten())
+            polars_emb = self.add_angle_proj(polars.flatten())
+            azimuths_emb = self.add_angle_proj(azimuths.flatten())
+            time_embeds = torch.cat([cond_aug_emb, polars_emb, azimuths_emb],dim=1)
+            time_embeds = time_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(time_embeds)
+            emb = emb + aug_emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+        # 3. down
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        is_adapter = mid_block_additional_residual is None and down_block_additional_residuals is not None
+
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_block_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_block_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+                if is_adapter and len(down_block_additional_residuals) > 0:
+                    sample += down_block_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            sample = self.mid_block(
+                sample,
+                emb,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=attention_mask,
+                cross_attention_kwargs=cross_attention_kwargs,
+                encoder_attention_mask=encoder_attention_mask,
+            )
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_block_additional_residuals) > 0
+                and sample.shape == down_block_additional_residuals[0].shape
+            ):
+                sample += down_block_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
+                )
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

ref_encoder/reference_unetv2.py

@@ -0,0 +1,1037 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.lora import LoRALinearLayer
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    PositionNet,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.unet_2d_blocks import (
+    UNetMidBlock2DCrossAttn,
+    UNetMidBlock2DSimpleCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+
+class CCProjection(ModelMixin, ConfigMixin):
+    def __init__(self, in_channel=772, out_channel=768):
+        super().__init__()
+        self.in_channel = in_channel
+        self.out_channel = out_channel
+        self.projection = torch.nn.Linear(in_channel, out_channel)
+
+    def forward(self, x):
+        return self.projection(x)
+
+
+class Identity(torch.nn.Module):
+    def __init__(self, scale=None, *args, **kwargs) -> None:
+        super(Identity, self).__init__()
+    def forward(self, input, *args, **kwargs):
+        return input
+
+
+class _LoRACompatibleLinear(nn.Module):
+    """
+    A Linear layer that can be used with LoRA.
+    """
+
+    def __init__(self, *args, lora_layer: Optional[LoRALinearLayer] = None, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.lora_layer = lora_layer
+
+    def set_lora_layer(self, lora_layer: Optional[LoRALinearLayer]):
+        self.lora_layer = lora_layer
+
+    def _fuse_lora(self):
+        pass
+
+    def _unfuse_lora(self):
+        pass
+
+    def forward(self, hidden_states, scale=None, lora_scale: int = 1):
+        return hidden_states
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+
+
+class ref_unet(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be either `UNetMidBlock2DCrossAttn` or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
+        conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
+        projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: int = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads=64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        # time
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                attention_type=attention_type,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                cross_attention_dim=cross_attention_dim[-1],
+                attention_head_dim=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                skip_time_act=resnet_skip_time_act,
+                only_cross_attention=mid_block_only_cross_attention,
+                cross_attention_norm=cross_attention_norm,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+        self.up_blocks[3].attentions[2].transformer_blocks[0].attn1.to_q = _LoRACompatibleLinear()
+        self.up_blocks[3].attentions[2].transformer_blocks[0].attn1.to_k = _LoRACompatibleLinear()
+        self.up_blocks[3].attentions[2].transformer_blocks[0].attn1.to_v = _LoRACompatibleLinear()
+        self.up_blocks[3].attentions[2].transformer_blocks[0].attn1.to_out = nn.ModuleList([Identity(), Identity()])
+        self.up_blocks[3].attentions[2].transformer_blocks[0].norm2 = Identity()
+        self.up_blocks[3].attentions[2].transformer_blocks[0].attn2 = None
+        self.up_blocks[3].attentions[2].transformer_blocks[0].norm3 = Identity()
+        self.up_blocks[3].attentions[2].transformer_blocks[0].ff = Identity()
+        self.up_blocks[3].attentions[2].proj_out = Identity()
+
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = PositionNet(
+                positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention face_hair_mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the face_hair_mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb, hint = self.add_embedding(image_embs, hint)
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+        # 3. down
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        is_adapter = mid_block_additional_residual is None and down_block_additional_residuals is not None
+
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_block_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_block_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+                if is_adapter and len(down_block_additional_residuals) > 0:
+                    sample += down_block_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            sample = self.mid_block(
+                sample,
+                emb,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=attention_mask,
+                cross_attention_kwargs=cross_attention_kwargs,
+                encoder_attention_mask=encoder_attention_mask,
+            )
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_block_additional_residuals) > 0
+                and sample.shape == down_block_additional_residuals[0].shape
+            ):
+                sample += down_block_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
+                )
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

ref_encoder/resnet.py

@@ -0,0 +1,263 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/resnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+
+class InflatedConv3d(nn.Conv2d):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class InflatedGroupNorm(nn.GroupNorm):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class Upsample3D(nn.Module):
+    def __init__(
+        self,
+        channels,
+        use_conv=False,
+        use_conv_transpose=False,
+        out_channels=None,
+        name="conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        conv = None
+        if use_conv_transpose:
+            raise NotImplementedError
+        elif use_conv:
+            self.conv = InflatedConv3d(self.channels, self.out_channels, 3, padding=1)
+
+    def forward(self, hidden_states, output_size=None):
+        assert hidden_states.shape[1] == self.channels
+
+        if self.use_conv_transpose:
+            raise NotImplementedError
+
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
+        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+        if hidden_states.shape[0] >= 64:
+            hidden_states = hidden_states.contiguous()
+
+        # if `output_size` is passed we force the interpolation output
+        # size and do not make use of `scale_factor=2`
+        if output_size is None:
+            hidden_states = F.interpolate(
+                hidden_states, scale_factor=[1.0, 2.0, 2.0], mode="nearest"
+            )
+        else:
+            hidden_states = F.interpolate(
+                hidden_states, size=output_size, mode="nearest"
+            )
+
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
+        # if self.use_conv:
+        #     if self.name == "conv":
+        #         hidden_states = self.conv(hidden_states)
+        #     else:
+        #         hidden_states = self.Conv2d_0(hidden_states)
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class Downsample3D(nn.Module):
+    def __init__(
+        self, channels, use_conv=False, out_channels=None, padding=1, name="conv"
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+
+        if use_conv:
+            self.conv = InflatedConv3d(
+                self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            raise NotImplementedError
+
+    def forward(self, hidden_states):
+        assert hidden_states.shape[1] == self.channels
+        if self.use_conv and self.padding == 0:
+            raise NotImplementedError
+
+        assert hidden_states.shape[1] == self.channels
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class ResnetBlock3D(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout=0.0,
+        temb_channels=512,
+        groups=32,
+        groups_out=None,
+        pre_norm=True,
+        eps=1e-6,
+        non_linearity="swish",
+        time_embedding_norm="default",
+        output_scale_factor=1.0,
+        use_in_shortcut=None,
+        use_inflated_groupnorm=None,
+    ):
+        super().__init__()
+        self.pre_norm = pre_norm
+        self.pre_norm = True
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.time_embedding_norm = time_embedding_norm
+        self.output_scale_factor = output_scale_factor
+
+        if groups_out is None:
+            groups_out = groups
+
+        assert use_inflated_groupnorm != None
+        if use_inflated_groupnorm:
+            self.norm1 = InflatedGroupNorm(
+                num_groups=groups, num_channels=in_channels, eps=eps, affine=True
+            )
+        else:
+            self.norm1 = torch.nn.GroupNorm(
+                num_groups=groups, num_channels=in_channels, eps=eps, affine=True
+            )
+
+        self.conv1 = InflatedConv3d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+        if temb_channels is not None:
+            if self.time_embedding_norm == "default":
+                time_emb_proj_out_channels = out_channels
+            elif self.time_embedding_norm == "scale_shift":
+                time_emb_proj_out_channels = out_channels * 2
+            else:
+                raise ValueError(
+                    f"unknown time_embedding_norm : {self.time_embedding_norm} "
+                )
+
+            self.time_emb_proj = torch.nn.Linear(
+                temb_channels, time_emb_proj_out_channels
+            )
+        else:
+            self.time_emb_proj = None
+
+        if use_inflated_groupnorm:
+            self.norm2 = InflatedGroupNorm(
+                num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
+            )
+        else:
+            self.norm2 = torch.nn.GroupNorm(
+                num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
+            )
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = InflatedConv3d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+        if non_linearity == "swish":
+            self.nonlinearity = lambda x: F.silu(x)
+        elif non_linearity == "mish":
+            self.nonlinearity = Mish()
+        elif non_linearity == "silu":
+            self.nonlinearity = nn.SiLU()
+
+        self.use_in_shortcut = (
+            self.in_channels != self.out_channels
+            if use_in_shortcut is None
+            else use_in_shortcut
+        )
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = InflatedConv3d(
+                in_channels, out_channels, kernel_size=1, stride=1, padding=0
+            )
+
+    def forward(self, input_tensor, temb):
+        hidden_states = input_tensor
+
+        video_length = input_tensor.shape[2]
+        batch_size = input_tensor.shape[0]
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv1(hidden_states)
+
+        if temb is not None:
+            temb_shape = temb.shape[0]
+            temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+            if temb_shape>batch_size:
+                temb = temb.permute(2, 1, 0, 3, 4)
+
+        
+        
+        if temb is not None and self.time_embedding_norm == "default":
+            if hidden_states.shape[0]>1 and temb.shape[2]==hidden_states.shape[0]*hidden_states.shape[2]:
+                temb = temb.reshape(temb.shape[0], temb.shape[1], hidden_states.shape[0], -1, temb.shape[-2], temb.shape[-1])
+                temb = temb.permute(2, 0, 1, 3, 4, 5)
+                temb = temb.reshape(temb.shape[0], -1, temb.shape[3], temb.shape[4], temb.shape[5])
+            hidden_states = hidden_states + temb
+
+        hidden_states = self.norm2(hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+class Mish(torch.nn.Module):
+    def forward(self, hidden_states):
+        return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))

ref_encoder/transformer_2d.py

@@ -0,0 +1,396 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformer_2d.py
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.embeddings import CaptionProjection
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
+from torch import nn
+
+from .attention import BasicTransformerBlock
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+    ref_feature: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate(
+                "norm_type!=num_embeds_ada_norm",
+                "1.0.0",
+                deprecation_message,
+                standard_warn=False,
+            )
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif (
+            not self.is_input_continuous
+            and not self.is_input_vectorized
+            and not self.is_input_patches
+        ):
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups,
+            num_channels=in_channels,
+            eps=1e-6,
+            affine=True,
+        )
+        if use_linear_projection:
+            self.proj_in = linear_cls(in_channels, inner_dim)
+        else:
+            self.proj_in = conv_cls(
+                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        # TODO: should use out_channels for continuous projections
+        if use_linear_projection:
+            self.proj_out = linear_cls(inner_dim, in_channels)
+        else:
+            self.proj_out = conv_cls(
+                inner_dim, in_channels, kernel_size=1, stride=1, padding=0
+            )
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(
+                inner_dim, use_additional_conditions=self.use_additional_conditions
+            )
+
+        self.caption_projection = None
+        if caption_channels is not None:
+            self.caption_projection = CaptionProjection(
+                in_features=caption_channels, hidden_size=inner_dim
+            )
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (
+                1 - encoder_attention_mask.to(hidden_states.dtype)
+            ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # Retrieve lora scale.
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        # 1. Input
+        batch, _, height, width = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale)
+                if not USE_PEFT_BACKEND
+                else self.proj_in(hidden_states)
+            )
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * width, inner_dim
+            )
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * width, inner_dim
+            )
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale)
+                if not USE_PEFT_BACKEND
+                else self.proj_in(hidden_states)
+            )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(
+                batch_size, -1, hidden_states.shape[-1]
+            )
+
+        ref_feature = hidden_states.reshape(batch, height, width, inner_dim)
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            if not self.use_linear_projection:
+                hidden_states = (
+                    hidden_states.reshape(batch, height, width, inner_dim)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+            else:
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+                hidden_states = (
+                    hidden_states.reshape(batch, height, width, inner_dim)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+
+            output = hidden_states + residual
+        if not return_dict:
+            return (output, ref_feature)
+
+        return Transformer2DModelOutput(sample=output, ref_feature=ref_feature)

ref_encoder/transformer_3d.py

@@ -0,0 +1,202 @@
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models import ModelMixin
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from einops import rearrange, repeat
+from torch import nn
+
+from .attention import TemporalBasicTransformerBlock
+
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        name=None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # Define input layers
+        self.in_channels = in_channels
+        self.name=name
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
+        )
+        if use_linear_projection:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_in = nn.Conv2d(
+                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+            )
+
+        # Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TemporalBasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                    name=f"{self.name}_{d}_TransformerBlock" if self.name else None,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        if use_linear_projection:
+            self.proj_out = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_out = nn.Conv2d(
+                inner_dim, in_channels, kernel_size=1, stride=1, padding=0
+            )
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        self_attention_additional_feats=None,
+        mode=None,
+        timestep=None,
+        return_dict: bool = True,
+    ):
+        # Input
+        assert (
+            hidden_states.dim() == 5
+        ), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+        if encoder_hidden_states.shape[0] != hidden_states.shape[0]:
+            encoder_hidden_states = repeat(
+                encoder_hidden_states, "b n c -> (b f) n c", f=video_length
+            )
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = self.proj_in(hidden_states)
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * weight, inner_dim
+            )
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * weight, inner_dim
+            )
+            hidden_states = self.proj_in(hidden_states)
+
+        # Blocks
+        for i, block in enumerate(self.transformer_blocks):
+            
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                # if hasattr(self.block, 'bank') and len(self.block.bank) > 0:
+                #     hidden_states
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    timestep=timestep,
+                    attention_mask=None,
+                    video_length=video_length,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=mode,
+                )
+            else:
+
+                hidden_states = block(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    timestep=timestep,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=mode,
+                    video_length=video_length,
+                )
+
+        # Output
+        if not self.use_linear_projection:
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim)
+                .permute(0, 3, 1, 2)
+                .contiguous()
+            )
+            hidden_states = self.proj_out(hidden_states)
+        else:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim)
+                .permute(0, 3, 1, 2)
+                .contiguous()
+            )
+
+        output = hidden_states + residual
+
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+        if not return_dict:
+            return (output,)
+
+        return Transformer3DModelOutput(sample=output)

ref_encoder/unet_2d_blocks.py

@@ -0,0 +1,1074 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+from typing import Any, Dict, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import Attention
+from diffusers.models.dual_transformer_2d import DualTransformer2DModel
+from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
+from diffusers.utils import is_torch_version, logging
+from diffusers.utils.torch_utils import apply_freeu
+from torch import nn
+
+from .transformer_2d import Transformer2DModel
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def get_down_block(
+    down_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    temb_channels: int,
+    add_downsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    downsample_padding: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    downsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+):
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    down_block_type = (
+        down_block_type[7:]
+        if down_block_type.startswith("UNetRes")
+        else down_block_type
+    )
+    if down_block_type == "DownBlock2D":
+        return DownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "CrossAttnDownBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnDownBlock2D"
+            )
+        return CrossAttnDownBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    prev_output_channel: int,
+    temb_channels: int,
+    add_upsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    resolution_idx: Optional[int] = None,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    upsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+) -> nn.Module:
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    up_block_type = (
+        up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    )
+    if up_block_type == "UpBlock2D":
+        return UpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif up_block_type == "CrossAttnUpBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnUpBlock2D"
+            )
+        return CrossAttnUpBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class AutoencoderTinyBlock(nn.Module):
+    """
+    Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
+    blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        out_channels (`int`): The number of output channels.
+        act_fn (`str`):
+            ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
+
+    Returns:
+        `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
+        `out_channels`.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, act_fn: str):
+        super().__init__()
+        act_fn = get_activation(act_fn)
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+        )
+        self.skip = (
+            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+        self.fuse = nn.ReLU()
+
+    def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        return self.fuse(self.conv(x) + self.skip(x))
+
+
+class UNetMidBlock2D(nn.Module):
+    """
+    A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        temb_channels (`int`): The number of temporal embedding channels.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
+        num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
+        resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
+        resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
+            The type of normalization to apply to the time embeddings. This can help to improve the performance of the
+            model on tasks with long-range temporal dependencies.
+        resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
+        resnet_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use in the group normalization layers of the resnet blocks.
+        attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
+        resnet_pre_norm (`bool`, *optional*, defaults to `True`):
+            Whether to use pre-normalization for the resnet blocks.
+        add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
+        attention_head_dim (`int`, *optional*, defaults to 1):
+            Dimension of a single attention head. The number of attention heads is determined based on this value and
+            the number of input channels.
+        output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
+
+    Returns:
+        `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
+        in_channels, height, width)`.
+
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",  # default, spatial
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        attn_groups: Optional[int] = None,
+        resnet_pre_norm: bool = True,
+        add_attention: bool = True,
+        attention_head_dim: int = 1,
+        output_scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+        self.add_attention = add_attention
+
+        if attn_groups is None:
+            attn_groups = (
+                resnet_groups if resnet_time_scale_shift == "default" else None
+            )
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        if attention_head_dim is None:
+            logger.warn(
+                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+            )
+            attention_head_dim = in_channels
+
+        for _ in range(num_layers):
+            if self.add_attention:
+                attentions.append(
+                    Attention(
+                        in_channels,
+                        heads=in_channels // attention_head_dim,
+                        dim_head=attention_head_dim,
+                        rescale_output_factor=output_scale_factor,
+                        eps=resnet_eps,
+                        norm_num_groups=attn_groups,
+                        spatial_norm_dim=temb_channels
+                        if resnet_time_scale_shift == "spatial"
+                        else None,
+                        residual_connection=True,
+                        bias=True,
+                        upcast_softmax=True,
+                        _from_deprecated_attn_block=True,
+                    )
+                )
+            else:
+                attentions.append(None)
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(
+        self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None
+    ) -> torch.FloatTensor:
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                hidden_states = attn(hidden_states, temb=temb)
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        output_scale_factor: float = 1.0,
+        cross_attention_dim: int = 1280,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+
+        # support for variable transformer layers per block
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for i in range(num_layers):
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+        return hidden_states
+
+
+class CrossAttnDownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        downsample_padding: int = 1,
+        add_downsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        additional_residuals: Optional[torch.FloatTensor] = None,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        blocks = list(zip(self.resnets, self.attentions))
+
+        for i, (resnet, attn) in enumerate(blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+
+            # apply additional residuals to the output of the last pair of resnet and attention blocks
+            if i == len(blocks) - 1 and additional_residuals is not None:
+                hidden_states = hidden_states + additional_residuals
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_downsample: bool = True,
+        downsample_padding: int = 1,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        scale: float = 1.0,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        upsample_size: Optional[int] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(
+                    hidden_states, upsample_size, scale=lora_scale
+                )
+
+        return hidden_states
+
+
+class UpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        upsample_size: Optional[int] = None,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        for resnet in self.resnets:
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
+
+        return hidden_states

ref_encoder/unet_2d_condition.py

@@ -0,0 +1,1308 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    PositionNet,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    deprecate,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+
+from .unet_2d_blocks import (
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+    ref_features: Tuple[torch.FloatTensor] = None
+
+
+class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+       reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+        *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+        *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = (
+            "UpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: int = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads=64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(
+            only_cross_attention
+        ) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if (
+            isinstance(transformer_layers_per_block, list)
+            and reverse_transformer_layers_per_block is None
+        ):
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError(
+                        "Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet."
+                    )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=conv_in_kernel,
+            padding=conv_in_padding,
+        )
+
+        # time
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(
+                    f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}."
+                )
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2,
+                set_W_to_weight=False,
+                log=False,
+                flip_sin_to_cos=flip_sin_to_cos,
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(
+                block_out_channels[0], flip_sin_to_cos, freq_shift
+            )
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info(
+                "encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined."
+            )
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(
+                timestep_input_dim, time_embed_dim, act_fn=act_fn
+            )
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim,
+                time_embed_dim,
+                num_heads=addition_embed_type_num_heads,
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim,
+                image_embed_dim=cross_attention_dim,
+                time_embed_dim=time_embed_dim,
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(
+                addition_time_embed_dim, flip_sin_to_cos, freq_shift
+            )
+            self.add_embedding = TimestepEmbedding(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(
+                image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(
+                image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type is not None:
+            raise ValueError(
+                f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'."
+            )
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(
+                down_block_types
+            )
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i]
+                if attention_head_dim[i] is not None
+                else output_channel,
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                attention_type=attention_type,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            raise NotImplementedError(f"Unsupport mid_block_type: {mid_block_type}")
+        elif mid_block_type == "UNetMidBlock2D":
+            self.mid_block = UNetMidBlock2D(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                num_layers=0,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                add_attention=False,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[
+                min(i + 1, len(block_out_channels) - 1)
+            ]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i]
+                if attention_head_dim[i] is not None
+                else output_channel,
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+        self.conv_norm_out = None
+
+        conv_out_padding = (conv_out_kernel - 1) // 2
+        # self.conv_out = nn.Conv2d(
+        #     block_out_channels[0],
+        #     out_channels,
+        #     kernel_size=conv_out_kernel,
+        #     padding=conv_out_padding,
+        # )
+
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(
+                cross_attention_dim, list
+            ):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = PositionNet(
+                positive_len=positive_len,
+                out_dim=cross_attention_dim,
+                feature_type=feature_type,
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(
+                    return_deprecated_lora=True
+                )
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(
+        self,
+        processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]],
+        _remove_lora=False,
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(
+                        processor.pop(f"{name}.processor"), _remove_lora=_remove_lora
+                    )
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(
+            proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS
+            for proc in self.attn_processors.values()
+        ):
+            processor = AttnAddedKVProcessor()
+        elif all(
+            proc.__class__ in CROSS_ATTENTION_PROCESSORS
+            for proc in self.attn_processors.values()
+        ):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = (
+            num_sliceable_layers * [slice_size]
+            if not isinstance(slice_size, list)
+            else slice_size
+        )
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(
+            module: torch.nn.Module, slice_size: List[int]
+        ):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1, s2, b1, b2):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if (
+                    hasattr(upsample_block, k)
+                    or getattr(upsample_block, k, None) is not None
+                ):
+                    setattr(upsample_block, k, None)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+                example from ControlNet side model(s)
+            mid_block_additional_residual (`torch.Tensor`, *optional*):
+                additional residual to be added to UNet mid block output, for example from ControlNet side model
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (
+                1 - encoder_attention_mask.to(sample.dtype)
+            ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError(
+                    "class_labels should be provided when num_class_embeds > 0"
+                )
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if (
+                "image_embeds" not in added_cond_kwargs
+                or "hint" not in added_cond_kwargs
+            ):
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb, hint = self.add_embedding(image_embs, hint)
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "text_proj"
+        ):
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "text_image_proj"
+        ):
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(
+                encoder_hidden_states, image_embeds
+            )
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "image_proj"
+        ):
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "ip_image_proj"
+        ):
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds).to(
+                encoder_hidden_states.dtype
+            )
+            encoder_hidden_states = torch.cat(
+                [encoder_hidden_states, image_embeds], dim=1
+            )
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if (
+            cross_attention_kwargs is not None
+            and cross_attention_kwargs.get("gligen", None) is not None
+        ):
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {
+                "objs": self.position_net(**gligen_args)
+            }
+
+        # 3. down
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = (
+            mid_block_additional_residual is not None
+            and down_block_additional_residuals is not None
+        )
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if (
+            not is_adapter
+            and mid_block_additional_residual is None
+            and down_block_additional_residuals is not None
+        ):
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        tot_referece_features = ()
+        for downsample_block in self.down_blocks:
+            if (
+                hasattr(downsample_block, "has_cross_attention")
+                and downsample_block.has_cross_attention
+            ):
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals[
+                        "additional_residuals"
+                    ] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample, temb=emb, scale=lora_scale
+                )
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = (
+                    down_block_res_sample + down_block_additional_residual
+                )
+                new_down_block_res_samples = new_down_block_res_samples + (
+                    down_block_res_sample,
+                )
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if (
+                hasattr(self.mid_block, "has_cross_attention")
+                and self.mid_block.has_cross_attention
+            ):
+                sample = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[
+                : -len(upsample_block.resnets)
+            ]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if (
+                hasattr(upsample_block, "has_cross_attention")
+                and upsample_block.has_cross_attention
+            ):
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    scale=lora_scale,
+                )
+
+        # 6. post-process
+        # if self.conv_norm_out:
+        #     sample = self.conv_norm_out(sample)
+        #     sample = self.conv_act(sample)
+        # sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

ref_encoder/unet_3d.py

@@ -0,0 +1,702 @@
+# Adapted from https://github.com/guoyww/AnimateDiff/blob/main/animatediff/models/unet_blocks.py
+
+from collections import OrderedDict
+from dataclasses import dataclass
+from os import PathLike
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.attention_processor import AttentionProcessor
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+try:
+    from diffusers.modeling_utils import ModelMixin
+except:
+    from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import SAFETENSORS_WEIGHTS_NAME, WEIGHTS_NAME, BaseOutput, logging
+from safetensors.torch import load_file
+
+from .resnet import InflatedConv3d, InflatedGroupNorm
+from .unet_3d_blocks import UNetMidBlock3DCrossAttn, get_down_block, get_up_block
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+from einops import rearrange
+
+
+@dataclass
+class UNet3DConditionOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+class UNet3DConditionModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ),
+        mid_block_type: str = "UNetMidBlock3DCrossAttn",
+        up_block_types: Tuple[str] = (
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: int = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: int = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: int = 1280,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        use_inflated_groupnorm=False,
+        # Additional
+        use_motion_module=False,  ######
+        motion_module_resolutions=(1, 2, 4, 8),  ####
+        motion_module_mid_block=False,   #####
+        motion_module_decoder_only=False,   #####
+        motion_module_type=None,  #####
+        motion_module_kwargs={},   #####
+        unet_use_cross_frame_attention=None,  #####
+        unet_use_temporal_attention=None,  #####
+        mode=None,  #####
+        task_type="action",   #####
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        time_embed_dim = block_out_channels[0] * 4
+
+        # input
+        self.conv_in = InflatedConv3d(
+            in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1)
+        )
+
+        # time
+        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+        timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        self.down_blocks = nn.ModuleList([])
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            if task_type == "action":
+                name_index, mid_name = None, None
+            else:
+                name_index, mid_name = i, "MidBlock"
+            res = 2**i
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module
+                and (res in motion_module_resolutions)
+                and (not motion_module_decoder_only),
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+                name_index=name_index,  #####
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        
+        if mid_block_type == "UNetMidBlock3DCrossAttn":
+            self.mid_block = UNetMidBlock3DCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module and motion_module_mid_block,
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+                name=mid_name,
+            )
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the videos
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_attention_head_dim = list(reversed(attention_head_dim))
+        only_cross_attention = list(reversed(only_cross_attention))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            res = 2 ** (3 - i)
+            is_final_block = i == len(block_out_channels) - 1
+            
+            if task_type == "action":
+                name_index = None
+            else:
+                name_index = i
+            
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[
+                min(i + 1, len(block_out_channels) - 1)
+            ]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=layers_per_block + 1,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=reversed_attention_head_dim[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module
+                and (res in motion_module_resolutions),
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+                name_index=name_index,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if use_inflated_groupnorm:
+            self.conv_norm_out = InflatedGroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+        else:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+        self.conv_act = nn.SiLU()
+        self.conv_out = InflatedConv3d(
+            block_out_channels[0], out_channels, kernel_size=3, padding=1
+        )
+        
+        self.mode = mode
+
+    @property
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "set_processor"):
+                processors[f"{name}.processor"] = module.processor
+
+            for sub_name, child in module.named_children():
+                if "temporal_transformer" not in sub_name:
+                    fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            if "temporal_transformer" not in name:
+                fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_slicable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_slicable_dims(module)
+
+        num_slicable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_slicable_layers * [1]
+
+        slice_size = (
+            num_slicable_layers * [slice_size]
+            if not isinstance(slice_size, list)
+            else slice_size
+        )
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(
+            module: torch.nn.Module, slice_size: List[int]
+        ):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                if "temporal_transformer" not in sub_name:
+                    fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            if "temporal_transformer" not in name:
+                fn_recursive_attn_processor(name, module, processor)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        pose_cond_fea: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        self_attention_additional_feats = None,
+    ) -> Union[UNet3DConditionOutput, Tuple]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
+            timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
+            encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=self.dtype)
+        emb = self.time_embedding(t_emb)
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError(
+                    "class_labels should be provided when num_class_embeds > 0"
+                )
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+
+        # pre-process
+        sample = self.conv_in(sample)
+        if pose_cond_fea is not None:
+            sample = sample + pose_cond_fea
+
+        # down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if (
+                hasattr(downsample_block, "has_cross_attention")
+                and downsample_block.has_cross_attention
+            ):
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=self.mode,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                )
+
+            down_block_res_samples += res_samples
+
+        if down_block_additional_residuals is not None:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_additional_residual = rearrange(
+                    down_block_additional_residual, "(b f) c h w -> b c f h w", f=down_block_res_sample.shape[2]
+                )
+                down_block_res_sample = (
+                    down_block_res_sample + down_block_additional_residual
+                )
+                new_down_block_res_samples += (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # mid
+        sample = self.mid_block(
+            sample,
+            emb,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            self_attention_additional_feats=self_attention_additional_feats,
+            mode=self.mode,
+        )
+
+        if mid_block_additional_residual is not None:
+            mid_block_additional_residual = rearrange(
+                mid_block_additional_residual, "(b f) c h w -> b c f h w", f=sample.shape[2]
+            )            
+            sample = sample + mid_block_additional_residual
+
+        # up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[
+                : -len(upsample_block.resnets)
+            ]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if (
+                hasattr(upsample_block, "has_cross_attention")
+                and upsample_block.has_cross_attention
+            ):
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=self.mode,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    encoder_hidden_states=encoder_hidden_states,
+                )
+
+        # post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet3DConditionOutput(sample=sample)
+
+    @classmethod
+    def from_pretrained_2d(
+        cls,
+        pretrained_model_path: PathLike,
+        motion_module_path: PathLike,
+        subfolder=None,
+        unet_additional_kwargs=None,
+        mm_zero_proj_out=False,
+    ):
+        pretrained_model_path = Path(pretrained_model_path)
+        motion_module_path = Path(motion_module_path)
+        if subfolder is not None:
+            pretrained_model_path = pretrained_model_path.joinpath(subfolder)
+        logger.info(
+            f"loaded temporal unet's pretrained weights from {pretrained_model_path} ..."
+        )
+
+        config_file = pretrained_model_path / "config.json"
+        if not (config_file.exists() and config_file.is_file()):
+            raise RuntimeError(f"{config_file} does not exist or is not a file")
+
+        unet_config = cls.load_config(config_file)
+        unet_config["_class_name"] = cls.__name__
+        unet_config["down_block_types"] = [
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ]
+        unet_config["up_block_types"] = [
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+        ]
+        unet_config["mid_block_type"] = "UNetMidBlock3DCrossAttn"
+
+        model = cls.from_config(unet_config, **unet_additional_kwargs)
+        # load the vanilla weights
+        if pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME).exists():
+            logger.debug(
+                f"loading safeTensors weights from {pretrained_model_path} ..."
+            )
+            state_dict = load_file(
+                pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME), device="cpu"
+            )
+
+        elif pretrained_model_path.joinpath(WEIGHTS_NAME).exists():
+            logger.debug(f"loading weights from {pretrained_model_path} ...")
+            state_dict = torch.load(
+                pretrained_model_path.joinpath(WEIGHTS_NAME),
+                map_location="cpu",
+                weights_only=True,
+            )
+        else:
+            raise FileNotFoundError(f"no weights file found in {pretrained_model_path}")
+
+        # load the motion module weights
+        if motion_module_path.exists() and motion_module_path.is_file():
+            if motion_module_path.suffix.lower() in [".pth", ".pt", ".ckpt"]:
+                logger.info(f"Load motion module params from {motion_module_path}")
+                motion_state_dict = torch.load(
+                    motion_module_path, map_location="cpu", weights_only=True
+                )
+            elif motion_module_path.suffix.lower() == ".safetensors":
+                motion_state_dict = load_file(motion_module_path, device="cpu")
+            else:
+                raise RuntimeError(
+                    f"unknown file format for motion module weights: {motion_module_path.suffix}"
+                )
+            if mm_zero_proj_out:
+                logger.info(f"Zero initialize proj_out layers in motion module...")
+                new_motion_state_dict = OrderedDict()
+                for k in motion_state_dict:
+                    if "proj_out" in k:
+                        continue
+                    new_motion_state_dict[k] = motion_state_dict[k]
+                motion_state_dict = new_motion_state_dict
+
+            # merge the state dicts
+            state_dict.update(motion_state_dict)
+
+        # load the weights into the model
+        m, u = model.load_state_dict(state_dict, strict=False)
+        logger.debug(f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
+
+        params = [
+            p.numel() if "temporal" in n else 0 for n, p in model.named_parameters()
+        ]
+        logger.info(f"Loaded {sum(params) / 1e6}M-parameter motion module")
+
+        return model

ref_encoder/unet_3d_blocks.py

@@ -0,0 +1,906 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+
+import pdb
+
+import torch
+from torch import nn
+
+from .motion_module import get_motion_module
+
+# from .motion_module import get_motion_module
+from .resnet import Downsample3D, ResnetBlock3D, Upsample3D
+from .transformer_3d import Transformer3DModel
+
+
+def get_down_block(
+    down_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    temb_channels,
+    add_downsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    downsample_padding=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    unet_use_cross_frame_attention=None,
+    unet_use_temporal_attention=None,
+    use_inflated_groupnorm=None,
+    use_motion_module=None,
+    motion_module_type=None,
+    motion_module_kwargs=None,
+    name_index=None,
+):
+    down_block_type = (
+        down_block_type[7:]
+        if down_block_type.startswith("UNetRes")
+        else down_block_type
+    )
+    if down_block_type == "DownBlock3D":
+        return DownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    elif down_block_type == "CrossAttnDownBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnDownBlock3D"
+            )
+        if name_index is not None:
+            name_index = f"CrossAttnDownBlock_{name_index}_"
+        return CrossAttnDownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+            unet_use_temporal_attention=unet_use_temporal_attention,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+            name=name_index,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    prev_output_channel,
+    temb_channels,
+    add_upsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    unet_use_cross_frame_attention=None,
+    unet_use_temporal_attention=None,
+    use_inflated_groupnorm=None,
+    use_motion_module=None,
+    motion_module_type=None,
+    motion_module_kwargs=None,
+    name_index=None,
+):
+    up_block_type = (
+        up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    )
+    if up_block_type == "UpBlock3D":
+        return UpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    elif up_block_type == "CrossAttnUpBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnUpBlock3D"
+            )
+        if name_index is not None:
+            name_index = f"CrossAttnUpBlock_{name_index}_"
+        return CrossAttnUpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+            unet_use_temporal_attention=unet_use_temporal_attention,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+            name=name_index,
+        )
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class UNetMidBlock3DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        output_scale_factor=1.0,
+        cross_attention_dim=1280,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+        name=None
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+        self.name = name
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock3D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+            )
+        ]
+        attentions = []
+        motion_modules = []
+        for i in range(num_layers):
+            if dual_cross_attention:
+                raise NotImplementedError
+            if self.name is not None:
+                attn_name = f"{self.name}_{i}_TransformerModel"
+            else:
+                attn_name = None
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    in_channels // attn_num_head_channels,
+                    in_channels=in_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                    name=attn_name,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=in_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+    def forward(
+        self,
+        hidden_states,
+        temb=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        self_attention_additional_feats=None,
+        mode=None,
+    ):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet, motion_module in zip(
+            self.attentions, self.resnets[1:], self.motion_modules
+        ):
+            hidden_states = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                self_attention_additional_feats=self_attention_additional_feats,
+                mode=mode,
+            ).sample
+            hidden_states = (
+                motion_module(
+                    hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                )
+                if motion_module is not None
+                else hidden_states
+            )
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class CrossAttnDownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        downsample_padding=1,
+        add_downsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+        name=None,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+        motion_modules = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        self.name=name
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            if self.name is not None:
+                attn_name = f"{self.name}_{i}_TransformerModel"
+            else:
+                attn_name = None
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                    name=attn_name,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        temb=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        self_attention_additional_feats=None,
+        mode=None,
+    ):
+        output_states = ()
+
+        for i, (resnet, attn, motion_module) in enumerate(
+            zip(self.resnets, self.attentions, self.motion_modules)
+        ):
+            # self.gradient_checkpointing = False
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                    self_attention_additional_feats,
+                    mode,
+                )[0]
+
+                # add motion module
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=mode,
+                ).sample
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_downsample=True,
+        downsample_padding=1,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        motion_modules = []
+
+        # use_motion_module = False
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
+        output_states = ()
+
+        for resnet, motion_module in zip(self.resnets, self.motion_modules):
+            # print(f"DownBlock3D {self.gradient_checkpointing = }")
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_motion_module=None,
+        use_inflated_groupnorm=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+        name=None
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+        motion_modules = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        self.name = name
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            if self.name is not None:
+                attn_name = f"{self.name}_{i}_TransformerModel"
+            else:
+                attn_name = None
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                    name=attn_name,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        res_hidden_states_tuple,
+        temb=None,
+        encoder_hidden_states=None,
+        upsample_size=None,
+        attention_mask=None,
+        self_attention_additional_feats=None,
+        mode=None,
+    ):
+        for i, (resnet, attn, motion_module) in enumerate(
+            zip(self.resnets, self.attentions, self.motion_modules)
+        ):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                    self_attention_additional_feats,
+                    mode,
+                )[0]
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=mode,
+                ).sample
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states
+
+
+class UpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        motion_modules = []
+
+        # use_motion_module = False
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        res_hidden_states_tuple,
+        temb=None,
+        upsample_size=None,
+        encoder_hidden_states=None,
+    ):
+        for resnet, motion_module in zip(self.resnets, self.motion_modules):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            # print(f"UpBlock3D {self.gradient_checkpointing = }")
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states

src/dataset/dance_image.py

@@ -0,0 +1,124 @@
+import json
+import random
+
+import torch
+import torchvision.transforms as transforms
+from decord import VideoReader
+from PIL import Image
+from torch.utils.data import Dataset
+from transformers import CLIPImageProcessor
+
+
+class HumanDanceDataset(Dataset):
+    def __init__(
+        self,
+        img_size,
+        img_scale=(1.0, 1.0),
+        img_ratio=(0.9, 1.0),
+        drop_ratio=0.1,
+        data_meta_paths=["./data/fahsion_meta.json"],
+        sample_margin=30,
+    ):
+        super().__init__()
+
+        self.img_size = img_size
+        self.img_scale = img_scale
+        self.img_ratio = img_ratio
+        self.sample_margin = sample_margin
+
+        # -----
+        # vid_meta format:
+        # [{'video_path': , 'kps_path': , 'other':},
+        #  {'video_path': , 'kps_path': , 'other':}]
+        # -----
+        vid_meta = []
+        for data_meta_path in data_meta_paths:
+            vid_meta.extend(json.load(open(data_meta_path, "r")))
+        self.vid_meta = vid_meta
+
+        self.clip_image_processor = CLIPImageProcessor()
+
+        self.transform = transforms.Compose(
+            [
+                transforms.RandomResizedCrop(
+                    self.img_size,
+                    scale=self.img_scale,
+                    ratio=self.img_ratio,
+                    interpolation=transforms.InterpolationMode.BILINEAR,
+                ),
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+
+        self.cond_transform = transforms.Compose(
+            [
+                transforms.RandomResizedCrop(
+                    self.img_size,
+                    scale=self.img_scale,
+                    ratio=self.img_ratio,
+                    interpolation=transforms.InterpolationMode.BILINEAR,
+                ),
+                transforms.ToTensor(),
+            ]
+        )
+
+        self.drop_ratio = drop_ratio
+
+    def augmentation(self, image, transform, state=None):
+        if state is not None:
+            torch.set_rng_state(state)
+        return transform(image)
+
+    def __getitem__(self, index):
+        video_meta = self.vid_meta[index]
+        video_path = video_meta["video_path"]
+        kps_path = video_meta["kps_path"]
+
+        video_reader = VideoReader(video_path)
+        kps_reader = VideoReader(kps_path)
+
+        assert len(video_reader) == len(
+            kps_reader
+        ), f"{len(video_reader) = } != {len(kps_reader) = } in {video_path}"
+
+        video_length = len(video_reader)
+
+        margin = min(self.sample_margin, video_length)
+
+        ref_img_idx = random.randint(0, video_length - 1)
+        if ref_img_idx + margin < video_length:
+            tgt_img_idx = random.randint(ref_img_idx + margin, video_length - 1)
+        elif ref_img_idx - margin > 0:
+            tgt_img_idx = random.randint(0, ref_img_idx - margin)
+        else:
+            tgt_img_idx = random.randint(0, video_length - 1)
+
+        ref_img = video_reader[ref_img_idx]
+        ref_img_pil = Image.fromarray(ref_img.asnumpy())
+        tgt_img = video_reader[tgt_img_idx]
+        tgt_img_pil = Image.fromarray(tgt_img.asnumpy())
+
+        tgt_pose = kps_reader[tgt_img_idx]
+        tgt_pose_pil = Image.fromarray(tgt_pose.asnumpy())
+
+        state = torch.get_rng_state()
+        tgt_img = self.augmentation(tgt_img_pil, self.transform, state)
+        tgt_pose_img = self.augmentation(tgt_pose_pil, self.cond_transform, state)
+        ref_img_vae = self.augmentation(ref_img_pil, self.transform, state)
+        clip_image = self.clip_image_processor(
+            images=ref_img_pil, return_tensors="pt"
+        ).pixel_values[0]
+
+        sample = dict(
+            video_dir=video_path,
+            img=tgt_img,
+            tgt_pose=tgt_pose_img,
+            ref_img=ref_img_vae,
+            clip_images=clip_image,
+        )
+
+        return sample
+
+    def __len__(self):
+        return len(self.vid_meta)

src/dataset/dance_video.py

@@ -0,0 +1,137 @@
+import json
+import random
+from typing import List
+
+import numpy as np
+import pandas as pd
+import torch
+import torchvision.transforms as transforms
+from decord import VideoReader
+from PIL import Image
+from torch.utils.data import Dataset
+from transformers import CLIPImageProcessor
+
+
+class HumanDanceVideoDataset(Dataset):
+    def __init__(
+        self,
+        sample_rate,
+        n_sample_frames,
+        width,
+        height,
+        img_scale=(1.0, 1.0),
+        img_ratio=(0.9, 1.0),
+        drop_ratio=0.1,
+        data_meta_paths=["./data/fashion_meta.json"],
+    ):
+        super().__init__()
+        self.sample_rate = sample_rate
+        self.n_sample_frames = n_sample_frames
+        self.width = width
+        self.height = height
+        self.img_scale = img_scale
+        self.img_ratio = img_ratio
+
+        vid_meta = []
+        for data_meta_path in data_meta_paths:
+            vid_meta.extend(json.load(open(data_meta_path, "r")))
+        self.vid_meta = vid_meta
+
+        self.clip_image_processor = CLIPImageProcessor()
+
+        self.pixel_transform = transforms.Compose(
+            [
+                transforms.RandomResizedCrop(
+                    (height, width),
+                    scale=self.img_scale,
+                    ratio=self.img_ratio,
+                    interpolation=transforms.InterpolationMode.BILINEAR,
+                ),
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+
+        self.cond_transform = transforms.Compose(
+            [
+                transforms.RandomResizedCrop(
+                    (height, width),
+                    scale=self.img_scale,
+                    ratio=self.img_ratio,
+                    interpolation=transforms.InterpolationMode.BILINEAR,
+                ),
+                transforms.ToTensor(),
+            ]
+        )
+
+        self.drop_ratio = drop_ratio
+
+    def augmentation(self, images, transform, state=None):
+        if state is not None:
+            torch.set_rng_state(state)
+        if isinstance(images, List):
+            transformed_images = [transform(img) for img in images]
+            ret_tensor = torch.stack(transformed_images, dim=0)  # (f, c, h, w)
+        else:
+            ret_tensor = transform(images)  # (c, h, w)
+        return ret_tensor
+
+    def __getitem__(self, index):
+        video_meta = self.vid_meta[index]
+        video_path = video_meta["video_path"]
+        kps_path = video_meta["kps_path"]
+
+        video_reader = VideoReader(video_path)
+        kps_reader = VideoReader(kps_path)
+
+        assert len(video_reader) == len(
+            kps_reader
+        ), f"{len(video_reader) = } != {len(kps_reader) = } in {video_path}"
+
+        video_length = len(video_reader)
+
+        clip_length = min(
+            video_length, (self.n_sample_frames - 1) * self.sample_rate + 1
+        )
+        start_idx = random.randint(0, video_length - clip_length)
+        batch_index = np.linspace(
+            start_idx, start_idx + clip_length - 1, self.n_sample_frames, dtype=int
+        ).tolist()
+
+        # read frames and kps
+        vid_pil_image_list = []
+        pose_pil_image_list = []
+        for index in batch_index:
+            img = video_reader[index]
+            vid_pil_image_list.append(Image.fromarray(img.asnumpy()))
+            img = kps_reader[index]
+            pose_pil_image_list.append(Image.fromarray(img.asnumpy()))
+
+        ref_img_idx = random.randint(0, video_length - 1)
+        ref_img = Image.fromarray(video_reader[ref_img_idx].asnumpy())
+
+        # transform
+        state = torch.get_rng_state()
+        pixel_values_vid = self.augmentation(
+            vid_pil_image_list, self.pixel_transform, state
+        )
+        pixel_values_pose = self.augmentation(
+            pose_pil_image_list, self.cond_transform, state
+        )
+        pixel_values_ref_img = self.augmentation(ref_img, self.pixel_transform, state)
+        clip_ref_img = self.clip_image_processor(
+            images=ref_img, return_tensors="pt"
+        ).pixel_values[0]
+
+        sample = dict(
+            video_dir=video_path,
+            pixel_values_vid=pixel_values_vid,
+            pixel_values_pose=pixel_values_pose,
+            pixel_values_ref_img=pixel_values_ref_img,
+            clip_ref_img=clip_ref_img,
+        )
+
+        return sample
+
+    def __len__(self):
+        return len(self.vid_meta)

src/dwpose/__init__.py

@@ -0,0 +1,123 @@
+# https://github.com/IDEA-Research/DWPose
+# Openpose
+# Original from CMU https://github.com/CMU-Perceptual-Computing-Lab/openpose
+# 2nd Edited by https://github.com/Hzzone/pytorch-openpose
+# 3rd Edited by ControlNet
+# 4th Edited by ControlNet (added face and correct hands)
+
+import copy
+import os
+
+os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
+import cv2
+import numpy as np
+import torch
+from controlnet_aux.util import HWC3, resize_image
+from PIL import Image
+
+from . import util
+from .wholebody import Wholebody
+
+
+def draw_pose(pose, H, W):
+    bodies = pose["bodies"]
+    faces = pose["faces"]
+    hands = pose["hands"]
+    candidate = bodies["candidate"]
+    subset = bodies["subset"]
+    canvas = np.zeros(shape=(H, W, 3), dtype=np.uint8)
+
+    canvas = util.draw_bodypose(canvas, candidate, subset)
+
+    canvas = util.draw_handpose(canvas, hands)
+
+    canvas = util.draw_facepose(canvas, faces)
+
+    return canvas
+
+
+class DWposeDetector:
+    def __init__(self):
+        pass
+
+    def to(self, device):
+        self.pose_estimation = Wholebody(device)
+        return self
+
+    def cal_height(self, input_image):
+        input_image = cv2.cvtColor(
+            np.array(input_image, dtype=np.uint8), cv2.COLOR_RGB2BGR
+        )
+
+        input_image = HWC3(input_image)
+        H, W, C = input_image.shape
+        with torch.no_grad():
+            candidate, subset = self.pose_estimation(input_image)
+            nums, keys, locs = candidate.shape
+            # candidate[..., 0] /= float(W)
+            # candidate[..., 1] /= float(H)
+            body = candidate
+        return body[0, ..., 1].min(), body[..., 1].max() - body[..., 1].min()
+
+    def __call__(
+        self,
+        input_image,
+        detect_resolution=512,
+        image_resolution=512,
+        output_type="pil",
+        **kwargs,
+    ):
+        input_image = cv2.cvtColor(
+            np.array(input_image, dtype=np.uint8), cv2.COLOR_RGB2BGR
+        )
+
+        input_image = HWC3(input_image)
+        input_image = resize_image(input_image, detect_resolution)
+        H, W, C = input_image.shape
+        with torch.no_grad():
+            candidate, subset = self.pose_estimation(input_image)
+            nums, keys, locs = candidate.shape
+            candidate[..., 0] /= float(W)
+            candidate[..., 1] /= float(H)
+            score = subset[:, :18]
+            max_ind = np.mean(score, axis=-1).argmax(axis=0)
+            score = score[[max_ind]]
+            body = candidate[:, :18].copy()
+            body = body[[max_ind]]
+            nums = 1
+            body = body.reshape(nums * 18, locs)
+            body_score = copy.deepcopy(score)
+            for i in range(len(score)):
+                for j in range(len(score[i])):
+                    if score[i][j] > 0.3:
+                        score[i][j] = int(18 * i + j)
+                    else:
+                        score[i][j] = -1
+
+            un_visible = subset < 0.3
+            candidate[un_visible] = -1
+
+            foot = candidate[:, 18:24]
+
+            faces = candidate[[max_ind], 24:92]
+
+            hands = candidate[[max_ind], 92:113]
+            hands = np.vstack([hands, candidate[[max_ind], 113:]])
+
+            bodies = dict(candidate=body, subset=score)
+            pose = dict(bodies=bodies, hands=hands, faces=faces)
+
+            detected_map = draw_pose(pose, H, W)
+            detected_map = HWC3(detected_map)
+
+            img = resize_image(input_image, image_resolution)
+            H, W, C = img.shape
+
+            detected_map = cv2.resize(
+                detected_map, (W, H), interpolation=cv2.INTER_LINEAR
+            )
+
+            if output_type == "pil":
+                detected_map = Image.fromarray(detected_map)
+
+            return detected_map, body_score

src/dwpose/onnxdet.py

@@ -0,0 +1,130 @@
+# https://github.com/IDEA-Research/DWPose
+import cv2
+import numpy as np
+import onnxruntime
+
+
+def nms(boxes, scores, nms_thr):
+    """Single class NMS implemented in Numpy."""
+    x1 = boxes[:, 0]
+    y1 = boxes[:, 1]
+    x2 = boxes[:, 2]
+    y2 = boxes[:, 3]
+
+    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+    order = scores.argsort()[::-1]
+
+    keep = []
+    while order.size > 0:
+        i = order[0]
+        keep.append(i)
+        xx1 = np.maximum(x1[i], x1[order[1:]])
+        yy1 = np.maximum(y1[i], y1[order[1:]])
+        xx2 = np.minimum(x2[i], x2[order[1:]])
+        yy2 = np.minimum(y2[i], y2[order[1:]])
+
+        w = np.maximum(0.0, xx2 - xx1 + 1)
+        h = np.maximum(0.0, yy2 - yy1 + 1)
+        inter = w * h
+        ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+        inds = np.where(ovr <= nms_thr)[0]
+        order = order[inds + 1]
+
+    return keep
+
+
+def multiclass_nms(boxes, scores, nms_thr, score_thr):
+    """Multiclass NMS implemented in Numpy. Class-aware version."""
+    final_dets = []
+    num_classes = scores.shape[1]
+    for cls_ind in range(num_classes):
+        cls_scores = scores[:, cls_ind]
+        valid_score_mask = cls_scores > score_thr
+        if valid_score_mask.sum() == 0:
+            continue
+        else:
+            valid_scores = cls_scores[valid_score_mask]
+            valid_boxes = boxes[valid_score_mask]
+            keep = nms(valid_boxes, valid_scores, nms_thr)
+            if len(keep) > 0:
+                cls_inds = np.ones((len(keep), 1)) * cls_ind
+                dets = np.concatenate(
+                    [valid_boxes[keep], valid_scores[keep, None], cls_inds], 1
+                )
+                final_dets.append(dets)
+    if len(final_dets) == 0:
+        return None
+    return np.concatenate(final_dets, 0)
+
+
+def demo_postprocess(outputs, img_size, p6=False):
+    grids = []
+    expanded_strides = []
+    strides = [8, 16, 32] if not p6 else [8, 16, 32, 64]
+
+    hsizes = [img_size[0] // stride for stride in strides]
+    wsizes = [img_size[1] // stride for stride in strides]
+
+    for hsize, wsize, stride in zip(hsizes, wsizes, strides):
+        xv, yv = np.meshgrid(np.arange(wsize), np.arange(hsize))
+        grid = np.stack((xv, yv), 2).reshape(1, -1, 2)
+        grids.append(grid)
+        shape = grid.shape[:2]
+        expanded_strides.append(np.full((*shape, 1), stride))
+
+    grids = np.concatenate(grids, 1)
+    expanded_strides = np.concatenate(expanded_strides, 1)
+    outputs[..., :2] = (outputs[..., :2] + grids) * expanded_strides
+    outputs[..., 2:4] = np.exp(outputs[..., 2:4]) * expanded_strides
+
+    return outputs
+
+
+def preprocess(img, input_size, swap=(2, 0, 1)):
+    if len(img.shape) == 3:
+        padded_img = np.ones((input_size[0], input_size[1], 3), dtype=np.uint8) * 114
+    else:
+        padded_img = np.ones(input_size, dtype=np.uint8) * 114
+
+    r = min(input_size[0] / img.shape[0], input_size[1] / img.shape[1])
+    resized_img = cv2.resize(
+        img,
+        (int(img.shape[1] * r), int(img.shape[0] * r)),
+        interpolation=cv2.INTER_LINEAR,
+    ).astype(np.uint8)
+    padded_img[: int(img.shape[0] * r), : int(img.shape[1] * r)] = resized_img
+
+    padded_img = padded_img.transpose(swap)
+    padded_img = np.ascontiguousarray(padded_img, dtype=np.float32)
+    return padded_img, r
+
+
+def inference_detector(session, oriImg):
+    input_shape = (640, 640)
+    img, ratio = preprocess(oriImg, input_shape)
+
+    ort_inputs = {session.get_inputs()[0].name: img[None, :, :, :]}
+    output = session.run(None, ort_inputs)
+    predictions = demo_postprocess(output[0], input_shape)[0]
+
+    boxes = predictions[:, :4]
+    scores = predictions[:, 4:5] * predictions[:, 5:]
+
+    boxes_xyxy = np.ones_like(boxes)
+    boxes_xyxy[:, 0] = boxes[:, 0] - boxes[:, 2] / 2.0
+    boxes_xyxy[:, 1] = boxes[:, 1] - boxes[:, 3] / 2.0
+    boxes_xyxy[:, 2] = boxes[:, 0] + boxes[:, 2] / 2.0
+    boxes_xyxy[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.0
+    boxes_xyxy /= ratio
+    dets = multiclass_nms(boxes_xyxy, scores, nms_thr=0.45, score_thr=0.1)
+    if dets is not None:
+        final_boxes, final_scores, final_cls_inds = dets[:, :4], dets[:, 4], dets[:, 5]
+        isscore = final_scores > 0.3
+        iscat = final_cls_inds == 0
+        isbbox = [i and j for (i, j) in zip(isscore, iscat)]
+        final_boxes = final_boxes[isbbox]
+    else:
+        return []
+
+    return final_boxes

src/dwpose/onnxpose.py

@@ -0,0 +1,370 @@
+# https://github.com/IDEA-Research/DWPose
+from typing import List, Tuple
+
+import cv2
+import numpy as np
+import onnxruntime as ort
+
+
+def preprocess(
+    img: np.ndarray, out_bbox, input_size: Tuple[int, int] = (192, 256)
+) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """Do preprocessing for RTMPose model inference.
+
+    Args:
+        img (np.ndarray): Input image in shape.
+        input_size (tuple): Input image size in shape (w, h).
+
+    Returns:
+        tuple:
+        - resized_img (np.ndarray): Preprocessed image.
+        - center (np.ndarray): Center of image.
+        - scale (np.ndarray): Scale of image.
+    """
+    # get shape of image
+    img_shape = img.shape[:2]
+    out_img, out_center, out_scale = [], [], []
+    if len(out_bbox) == 0:
+        out_bbox = [[0, 0, img_shape[1], img_shape[0]]]
+    for i in range(len(out_bbox)):
+        x0 = out_bbox[i][0]
+        y0 = out_bbox[i][1]
+        x1 = out_bbox[i][2]
+        y1 = out_bbox[i][3]
+        bbox = np.array([x0, y0, x1, y1])
+
+        # get center and scale
+        center, scale = bbox_xyxy2cs(bbox, padding=1.25)
+
+        # do affine transformation
+        resized_img, scale = top_down_affine(input_size, scale, center, img)
+
+        # normalize image
+        mean = np.array([123.675, 116.28, 103.53])
+        std = np.array([58.395, 57.12, 57.375])
+        resized_img = (resized_img - mean) / std
+
+        out_img.append(resized_img)
+        out_center.append(center)
+        out_scale.append(scale)
+
+    return out_img, out_center, out_scale
+
+
+def inference(sess: ort.InferenceSession, img: np.ndarray) -> np.ndarray:
+    """Inference RTMPose model.
+
+    Args:
+        sess (ort.InferenceSession): ONNXRuntime session.
+        img (np.ndarray): Input image in shape.
+
+    Returns:
+        outputs (np.ndarray): Output of RTMPose model.
+    """
+    all_out = []
+    # build input
+    for i in range(len(img)):
+        input = [img[i].transpose(2, 0, 1)]
+
+        # build output
+        sess_input = {sess.get_inputs()[0].name: input}
+        sess_output = []
+        for out in sess.get_outputs():
+            sess_output.append(out.name)
+
+        # run model
+        outputs = sess.run(sess_output, sess_input)
+        all_out.append(outputs)
+
+    return all_out
+
+
+def postprocess(
+    outputs: List[np.ndarray],
+    model_input_size: Tuple[int, int],
+    center: Tuple[int, int],
+    scale: Tuple[int, int],
+    simcc_split_ratio: float = 2.0,
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Postprocess for RTMPose model output.
+
+    Args:
+        outputs (np.ndarray): Output of RTMPose model.
+        model_input_size (tuple): RTMPose model Input image size.
+        center (tuple): Center of bbox in shape (x, y).
+        scale (tuple): Scale of bbox in shape (w, h).
+        simcc_split_ratio (float): Split ratio of simcc.
+
+    Returns:
+        tuple:
+        - keypoints (np.ndarray): Rescaled keypoints.
+        - scores (np.ndarray): Model predict scores.
+    """
+    all_key = []
+    all_score = []
+    for i in range(len(outputs)):
+        # use simcc to decode
+        simcc_x, simcc_y = outputs[i]
+        keypoints, scores = decode(simcc_x, simcc_y, simcc_split_ratio)
+
+        # rescale keypoints
+        keypoints = keypoints / model_input_size * scale[i] + center[i] - scale[i] / 2
+        all_key.append(keypoints[0])
+        all_score.append(scores[0])
+
+    return np.array(all_key), np.array(all_score)
+
+
+def bbox_xyxy2cs(
+    bbox: np.ndarray, padding: float = 1.0
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Transform the bbox format from (x,y,w,h) into (center, scale)
+
+    Args:
+        bbox (ndarray): Bounding box(es) in shape (4,) or (n, 4), formatted
+            as (left, top, right, bottom)
+        padding (float): BBox padding factor that will be multilied to scale.
+            Default: 1.0
+
+    Returns:
+        tuple: A tuple containing center and scale.
+        - np.ndarray[float32]: Center (x, y) of the bbox in shape (2,) or
+            (n, 2)
+        - np.ndarray[float32]: Scale (w, h) of the bbox in shape (2,) or
+            (n, 2)
+    """
+    # convert single bbox from (4, ) to (1, 4)
+    dim = bbox.ndim
+    if dim == 1:
+        bbox = bbox[None, :]
+
+    # get bbox center and scale
+    x1, y1, x2, y2 = np.hsplit(bbox, [1, 2, 3])
+    center = np.hstack([x1 + x2, y1 + y2]) * 0.5
+    scale = np.hstack([x2 - x1, y2 - y1]) * padding
+
+    if dim == 1:
+        center = center[0]
+        scale = scale[0]
+
+    return center, scale
+
+
+def _fix_aspect_ratio(bbox_scale: np.ndarray, aspect_ratio: float) -> np.ndarray:
+    """Extend the scale to match the given aspect ratio.
+
+    Args:
+        scale (np.ndarray): The image scale (w, h) in shape (2, )
+        aspect_ratio (float): The ratio of ``w/h``
+
+    Returns:
+        np.ndarray: The reshaped image scale in (2, )
+    """
+    w, h = np.hsplit(bbox_scale, [1])
+    bbox_scale = np.where(
+        w > h * aspect_ratio,
+        np.hstack([w, w / aspect_ratio]),
+        np.hstack([h * aspect_ratio, h]),
+    )
+    return bbox_scale
+
+
+def _rotate_point(pt: np.ndarray, angle_rad: float) -> np.ndarray:
+    """Rotate a point by an angle.
+
+    Args:
+        pt (np.ndarray): 2D point coordinates (x, y) in shape (2, )
+        angle_rad (float): rotation angle in radian
+
+    Returns:
+        np.ndarray: Rotated point in shape (2, )
+    """
+    sn, cs = np.sin(angle_rad), np.cos(angle_rad)
+    rot_mat = np.array([[cs, -sn], [sn, cs]])
+    return rot_mat @ pt
+
+
+def _get_3rd_point(a: np.ndarray, b: np.ndarray) -> np.ndarray:
+    """To calculate the affine matrix, three pairs of points are required. This
+    function is used to get the 3rd point, given 2D points a & b.
+
+    The 3rd point is defined by rotating vector `a - b` by 90 degrees
+    anticlockwise, using b as the rotation center.
+
+    Args:
+        a (np.ndarray): The 1st point (x,y) in shape (2, )
+        b (np.ndarray): The 2nd point (x,y) in shape (2, )
+
+    Returns:
+        np.ndarray: The 3rd point.
+    """
+    direction = a - b
+    c = b + np.r_[-direction[1], direction[0]]
+    return c
+
+
+def get_warp_matrix(
+    center: np.ndarray,
+    scale: np.ndarray,
+    rot: float,
+    output_size: Tuple[int, int],
+    shift: Tuple[float, float] = (0.0, 0.0),
+    inv: bool = False,
+) -> np.ndarray:
+    """Calculate the affine transformation matrix that can warp the bbox area
+    in the input image to the output size.
+
+    Args:
+        center (np.ndarray[2, ]): Center of the bounding box (x, y).
+        scale (np.ndarray[2, ]): Scale of the bounding box
+            wrt [width, height].
+        rot (float): Rotation angle (degree).
+        output_size (np.ndarray[2, ] | list(2,)): Size of the
+            destination heatmaps.
+        shift (0-100%): Shift translation ratio wrt the width/height.
+            Default (0., 0.).
+        inv (bool): Option to inverse the affine transform direction.
+            (inv=False: src->dst or inv=True: dst->src)
+
+    Returns:
+        np.ndarray: A 2x3 transformation matrix
+    """
+    shift = np.array(shift)
+    src_w = scale[0]
+    dst_w = output_size[0]
+    dst_h = output_size[1]
+
+    # compute transformation matrix
+    rot_rad = np.deg2rad(rot)
+    src_dir = _rotate_point(np.array([0.0, src_w * -0.5]), rot_rad)
+    dst_dir = np.array([0.0, dst_w * -0.5])
+
+    # get four corners of the src rectangle in the original image
+    src = np.zeros((3, 2), dtype=np.float32)
+    src[0, :] = center + scale * shift
+    src[1, :] = center + src_dir + scale * shift
+    src[2, :] = _get_3rd_point(src[0, :], src[1, :])
+
+    # get four corners of the dst rectangle in the input image
+    dst = np.zeros((3, 2), dtype=np.float32)
+    dst[0, :] = [dst_w * 0.5, dst_h * 0.5]
+    dst[1, :] = np.array([dst_w * 0.5, dst_h * 0.5]) + dst_dir
+    dst[2, :] = _get_3rd_point(dst[0, :], dst[1, :])
+
+    if inv:
+        warp_mat = cv2.getAffineTransform(np.float32(dst), np.float32(src))
+    else:
+        warp_mat = cv2.getAffineTransform(np.float32(src), np.float32(dst))
+
+    return warp_mat
+
+
+def top_down_affine(
+    input_size: dict, bbox_scale: dict, bbox_center: dict, img: np.ndarray
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Get the bbox image as the model input by affine transform.
+
+    Args:
+        input_size (dict): The input size of the model.
+        bbox_scale (dict): The bbox scale of the img.
+        bbox_center (dict): The bbox center of the img.
+        img (np.ndarray): The original image.
+
+    Returns:
+        tuple: A tuple containing center and scale.
+        - np.ndarray[float32]: img after affine transform.
+        - np.ndarray[float32]: bbox scale after affine transform.
+    """
+    w, h = input_size
+    warp_size = (int(w), int(h))
+
+    # reshape bbox to fixed aspect ratio
+    bbox_scale = _fix_aspect_ratio(bbox_scale, aspect_ratio=w / h)
+
+    # get the affine matrix
+    center = bbox_center
+    scale = bbox_scale
+    rot = 0
+    warp_mat = get_warp_matrix(center, scale, rot, output_size=(w, h))
+
+    # do affine transform
+    img = cv2.warpAffine(img, warp_mat, warp_size, flags=cv2.INTER_LINEAR)
+
+    return img, bbox_scale
+
+
+def get_simcc_maximum(
+    simcc_x: np.ndarray, simcc_y: np.ndarray
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Get maximum response location and value from simcc representations.
+
+    Note:
+        instance number: N
+        num_keypoints: K
+        heatmap height: H
+        heatmap width: W
+
+    Args:
+        simcc_x (np.ndarray): x-axis SimCC in shape (K, Wx) or (N, K, Wx)
+        simcc_y (np.ndarray): y-axis SimCC in shape (K, Wy) or (N, K, Wy)
+
+    Returns:
+        tuple:
+        - locs (np.ndarray): locations of maximum heatmap responses in shape
+            (K, 2) or (N, K, 2)
+        - vals (np.ndarray): values of maximum heatmap responses in shape
+            (K,) or (N, K)
+    """
+    N, K, Wx = simcc_x.shape
+    simcc_x = simcc_x.reshape(N * K, -1)
+    simcc_y = simcc_y.reshape(N * K, -1)
+
+    # get maximum value locations
+    x_locs = np.argmax(simcc_x, axis=1)
+    y_locs = np.argmax(simcc_y, axis=1)
+    locs = np.stack((x_locs, y_locs), axis=-1).astype(np.float32)
+    max_val_x = np.amax(simcc_x, axis=1)
+    max_val_y = np.amax(simcc_y, axis=1)
+
+    # get maximum value across x and y axis
+    mask = max_val_x > max_val_y
+    max_val_x[mask] = max_val_y[mask]
+    vals = max_val_x
+    locs[vals <= 0.0] = -1
+
+    # reshape
+    locs = locs.reshape(N, K, 2)
+    vals = vals.reshape(N, K)
+
+    return locs, vals
+
+
+def decode(
+    simcc_x: np.ndarray, simcc_y: np.ndarray, simcc_split_ratio
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Modulate simcc distribution with Gaussian.
+
+    Args:
+        simcc_x (np.ndarray[K, Wx]): model predicted simcc in x.
+        simcc_y (np.ndarray[K, Wy]): model predicted simcc in y.
+        simcc_split_ratio (int): The split ratio of simcc.
+
+    Returns:
+        tuple: A tuple containing center and scale.
+        - np.ndarray[float32]: keypoints in shape (K, 2) or (n, K, 2)
+        - np.ndarray[float32]: scores in shape (K,) or (n, K)
+    """
+    keypoints, scores = get_simcc_maximum(simcc_x, simcc_y)
+    keypoints /= simcc_split_ratio
+
+    return keypoints, scores
+
+
+def inference_pose(session, out_bbox, oriImg):
+    h, w = session.get_inputs()[0].shape[2:]
+    model_input_size = (w, h)
+    resized_img, center, scale = preprocess(oriImg, out_bbox, model_input_size)
+    outputs = inference(session, resized_img)
+    keypoints, scores = postprocess(outputs, model_input_size, center, scale)
+
+    return keypoints, scores

src/dwpose/util.py

@@ -0,0 +1,378 @@
+# https://github.com/IDEA-Research/DWPose
+import math
+import numpy as np
+import matplotlib
+import cv2
+
+
+eps = 0.01
+
+
+def smart_resize(x, s):
+    Ht, Wt = s
+    if x.ndim == 2:
+        Ho, Wo = x.shape
+        Co = 1
+    else:
+        Ho, Wo, Co = x.shape
+    if Co == 3 or Co == 1:
+        k = float(Ht + Wt) / float(Ho + Wo)
+        return cv2.resize(
+            x,
+            (int(Wt), int(Ht)),
+            interpolation=cv2.INTER_AREA if k < 1 else cv2.INTER_LANCZOS4,
+        )
+    else:
+        return np.stack([smart_resize(x[:, :, i], s) for i in range(Co)], axis=2)
+
+
+def smart_resize_k(x, fx, fy):
+    if x.ndim == 2:
+        Ho, Wo = x.shape
+        Co = 1
+    else:
+        Ho, Wo, Co = x.shape
+    Ht, Wt = Ho * fy, Wo * fx
+    if Co == 3 or Co == 1:
+        k = float(Ht + Wt) / float(Ho + Wo)
+        return cv2.resize(
+            x,
+            (int(Wt), int(Ht)),
+            interpolation=cv2.INTER_AREA if k < 1 else cv2.INTER_LANCZOS4,
+        )
+    else:
+        return np.stack([smart_resize_k(x[:, :, i], fx, fy) for i in range(Co)], axis=2)
+
+
+def padRightDownCorner(img, stride, padValue):
+    h = img.shape[0]
+    w = img.shape[1]
+
+    pad = 4 * [None]
+    pad[0] = 0  # up
+    pad[1] = 0  # left
+    pad[2] = 0 if (h % stride == 0) else stride - (h % stride)  # down
+    pad[3] = 0 if (w % stride == 0) else stride - (w % stride)  # right
+
+    img_padded = img
+    pad_up = np.tile(img_padded[0:1, :, :] * 0 + padValue, (pad[0], 1, 1))
+    img_padded = np.concatenate((pad_up, img_padded), axis=0)
+    pad_left = np.tile(img_padded[:, 0:1, :] * 0 + padValue, (1, pad[1], 1))
+    img_padded = np.concatenate((pad_left, img_padded), axis=1)
+    pad_down = np.tile(img_padded[-2:-1, :, :] * 0 + padValue, (pad[2], 1, 1))
+    img_padded = np.concatenate((img_padded, pad_down), axis=0)
+    pad_right = np.tile(img_padded[:, -2:-1, :] * 0 + padValue, (1, pad[3], 1))
+    img_padded = np.concatenate((img_padded, pad_right), axis=1)
+
+    return img_padded, pad
+
+
+def transfer(model, model_weights):
+    transfered_model_weights = {}
+    for weights_name in model.state_dict().keys():
+        transfered_model_weights[weights_name] = model_weights[
+            ".".join(weights_name.split(".")[1:])
+        ]
+    return transfered_model_weights
+
+
+def draw_bodypose(canvas, candidate, subset):
+    H, W, C = canvas.shape
+    candidate = np.array(candidate)
+    subset = np.array(subset)
+
+    stickwidth = 4
+
+    limbSeq = [
+        [2, 3],
+        [2, 6],
+        [3, 4],
+        [4, 5],
+        [6, 7],
+        [7, 8],
+        [2, 9],
+        [9, 10],
+        [10, 11],
+        [2, 12],
+        [12, 13],
+        [13, 14],
+        [2, 1],
+        [1, 15],
+        [15, 17],
+        [1, 16],
+        [16, 18],
+        [3, 17],
+        [6, 18],
+    ]
+
+    colors = [
+        [255, 0, 0],
+        [255, 85, 0],
+        [255, 170, 0],
+        [255, 255, 0],
+        [170, 255, 0],
+        [85, 255, 0],
+        [0, 255, 0],
+        [0, 255, 85],
+        [0, 255, 170],
+        [0, 255, 255],
+        [0, 170, 255],
+        [0, 85, 255],
+        [0, 0, 255],
+        [85, 0, 255],
+        [170, 0, 255],
+        [255, 0, 255],
+        [255, 0, 170],
+        [255, 0, 85],
+    ]
+
+    for i in range(17):
+        for n in range(len(subset)):
+            index = subset[n][np.array(limbSeq[i]) - 1]
+            if -1 in index:
+                continue
+            Y = candidate[index.astype(int), 0] * float(W)
+            X = candidate[index.astype(int), 1] * float(H)
+            mX = np.mean(X)
+            mY = np.mean(Y)
+            length = ((X[0] - X[1]) ** 2 + (Y[0] - Y[1]) ** 2) ** 0.5
+            angle = math.degrees(math.atan2(X[0] - X[1], Y[0] - Y[1]))
+            polygon = cv2.ellipse2Poly(
+                (int(mY), int(mX)), (int(length / 2), stickwidth), int(angle), 0, 360, 1
+            )
+            cv2.fillConvexPoly(canvas, polygon, colors[i])
+
+    canvas = (canvas * 0.6).astype(np.uint8)
+
+    for i in range(18):
+        for n in range(len(subset)):
+            index = int(subset[n][i])
+            if index == -1:
+                continue
+            x, y = candidate[index][0:2]
+            x = int(x * W)
+            y = int(y * H)
+            cv2.circle(canvas, (int(x), int(y)), 4, colors[i], thickness=-1)
+
+    return canvas
+
+
+def draw_handpose(canvas, all_hand_peaks):
+    H, W, C = canvas.shape
+
+    edges = [
+        [0, 1],
+        [1, 2],
+        [2, 3],
+        [3, 4],
+        [0, 5],
+        [5, 6],
+        [6, 7],
+        [7, 8],
+        [0, 9],
+        [9, 10],
+        [10, 11],
+        [11, 12],
+        [0, 13],
+        [13, 14],
+        [14, 15],
+        [15, 16],
+        [0, 17],
+        [17, 18],
+        [18, 19],
+        [19, 20],
+    ]
+
+    for peaks in all_hand_peaks:
+        peaks = np.array(peaks)
+
+        for ie, e in enumerate(edges):
+            x1, y1 = peaks[e[0]]
+            x2, y2 = peaks[e[1]]
+            x1 = int(x1 * W)
+            y1 = int(y1 * H)
+            x2 = int(x2 * W)
+            y2 = int(y2 * H)
+            if x1 > eps and y1 > eps and x2 > eps and y2 > eps:
+                cv2.line(
+                    canvas,
+                    (x1, y1),
+                    (x2, y2),
+                    matplotlib.colors.hsv_to_rgb([ie / float(len(edges)), 1.0, 1.0])
+                    * 255,
+                    thickness=2,
+                )
+
+        for i, keyponit in enumerate(peaks):
+            x, y = keyponit
+            x = int(x * W)
+            y = int(y * H)
+            if x > eps and y > eps:
+                cv2.circle(canvas, (x, y), 4, (0, 0, 255), thickness=-1)
+    return canvas
+
+
+def draw_facepose(canvas, all_lmks):
+    H, W, C = canvas.shape
+    for lmks in all_lmks:
+        lmks = np.array(lmks)
+        for lmk in lmks:
+            x, y = lmk
+            x = int(x * W)
+            y = int(y * H)
+            if x > eps and y > eps:
+                cv2.circle(canvas, (x, y), 3, (255, 255, 255), thickness=-1)
+    return canvas
+
+
+# detect hand according to body pose keypoints
+# please refer to https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/hand/handDetector.cpp
+def handDetect(candidate, subset, oriImg):
+    # right hand: wrist 4, elbow 3, shoulder 2
+    # left hand: wrist 7, elbow 6, shoulder 5
+    ratioWristElbow = 0.33
+    detect_result = []
+    image_height, image_width = oriImg.shape[0:2]
+    for person in subset.astype(int):
+        # if any of three not detected
+        has_left = np.sum(person[[5, 6, 7]] == -1) == 0
+        has_right = np.sum(person[[2, 3, 4]] == -1) == 0
+        if not (has_left or has_right):
+            continue
+        hands = []
+        # left hand
+        if has_left:
+            left_shoulder_index, left_elbow_index, left_wrist_index = person[[5, 6, 7]]
+            x1, y1 = candidate[left_shoulder_index][:2]
+            x2, y2 = candidate[left_elbow_index][:2]
+            x3, y3 = candidate[left_wrist_index][:2]
+            hands.append([x1, y1, x2, y2, x3, y3, True])
+        # right hand
+        if has_right:
+            right_shoulder_index, right_elbow_index, right_wrist_index = person[
+                [2, 3, 4]
+            ]
+            x1, y1 = candidate[right_shoulder_index][:2]
+            x2, y2 = candidate[right_elbow_index][:2]
+            x3, y3 = candidate[right_wrist_index][:2]
+            hands.append([x1, y1, x2, y2, x3, y3, False])
+
+        for x1, y1, x2, y2, x3, y3, is_left in hands:
+            # pos_hand = pos_wrist + ratio * (pos_wrist - pos_elbox) = (1 + ratio) * pos_wrist - ratio * pos_elbox
+            # handRectangle.x = posePtr[wrist*3] + ratioWristElbow * (posePtr[wrist*3] - posePtr[elbow*3]);
+            # handRectangle.y = posePtr[wrist*3+1] + ratioWristElbow * (posePtr[wrist*3+1] - posePtr[elbow*3+1]);
+            # const auto distanceWristElbow = getDistance(poseKeypoints, person, wrist, elbow);
+            # const auto distanceElbowShoulder = getDistance(poseKeypoints, person, elbow, shoulder);
+            # handRectangle.width = 1.5f * fastMax(distanceWristElbow, 0.9f * distanceElbowShoulder);
+            x = x3 + ratioWristElbow * (x3 - x2)
+            y = y3 + ratioWristElbow * (y3 - y2)
+            distanceWristElbow = math.sqrt((x3 - x2) ** 2 + (y3 - y2) ** 2)
+            distanceElbowShoulder = math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)
+            width = 1.5 * max(distanceWristElbow, 0.9 * distanceElbowShoulder)
+            # x-y refers to the center --> offset to topLeft point
+            # handRectangle.x -= handRectangle.width / 2.f;
+            # handRectangle.y -= handRectangle.height / 2.f;
+            x -= width / 2
+            y -= width / 2  # width = height
+            # overflow the image
+            if x < 0:
+                x = 0
+            if y < 0:
+                y = 0
+            width1 = width
+            width2 = width
+            if x + width > image_width:
+                width1 = image_width - x
+            if y + width > image_height:
+                width2 = image_height - y
+            width = min(width1, width2)
+            # the max hand box value is 20 pixels
+            if width >= 20:
+                detect_result.append([int(x), int(y), int(width), is_left])
+
+    """
+    return value: [[x, y, w, True if left hand else False]].
+    width=height since the network require squared input.
+    x, y is the coordinate of top left 
+    """
+    return detect_result
+
+
+# Written by Lvmin
+def faceDetect(candidate, subset, oriImg):
+    # left right eye ear 14 15 16 17
+    detect_result = []
+    image_height, image_width = oriImg.shape[0:2]
+    for person in subset.astype(int):
+        has_head = person[0] > -1
+        if not has_head:
+            continue
+
+        has_left_eye = person[14] > -1
+        has_right_eye = person[15] > -1
+        has_left_ear = person[16] > -1
+        has_right_ear = person[17] > -1
+
+        if not (has_left_eye or has_right_eye or has_left_ear or has_right_ear):
+            continue
+
+        head, left_eye, right_eye, left_ear, right_ear = person[[0, 14, 15, 16, 17]]
+
+        width = 0.0
+        x0, y0 = candidate[head][:2]
+
+        if has_left_eye:
+            x1, y1 = candidate[left_eye][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 3.0)
+
+        if has_right_eye:
+            x1, y1 = candidate[right_eye][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 3.0)
+
+        if has_left_ear:
+            x1, y1 = candidate[left_ear][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 1.5)
+
+        if has_right_ear:
+            x1, y1 = candidate[right_ear][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 1.5)
+
+        x, y = x0, y0
+
+        x -= width
+        y -= width
+
+        if x < 0:
+            x = 0
+
+        if y < 0:
+            y = 0
+
+        width1 = width * 2
+        width2 = width * 2
+
+        if x + width > image_width:
+            width1 = image_width - x
+
+        if y + width > image_height:
+            width2 = image_height - y
+
+        width = min(width1, width2)
+
+        if width >= 20:
+            detect_result.append([int(x), int(y), int(width)])
+
+    return detect_result
+
+
+# get max index of 2d array
+def npmax(array):
+    arrayindex = array.argmax(1)
+    arrayvalue = array.max(1)
+    i = arrayvalue.argmax()
+    j = arrayindex[i]
+    return i, j

src/dwpose/wholebody.py

@@ -0,0 +1,48 @@
+# https://github.com/IDEA-Research/DWPose
+from pathlib import Path
+
+import cv2
+import numpy as np
+import onnxruntime as ort
+
+from .onnxdet import inference_detector
+from .onnxpose import inference_pose
+
+ModelDataPathPrefix = Path("./pretrained_weights")
+
+
+class Wholebody:
+    def __init__(self, device="cuda:0"):
+        providers = (
+            ["CPUExecutionProvider"] if device == "cpu" else ["CUDAExecutionProvider"]
+        )
+        onnx_det = ModelDataPathPrefix.joinpath("DWPose/yolox_l.onnx")
+        onnx_pose = ModelDataPathPrefix.joinpath("DWPose/dw-ll_ucoco_384.onnx")
+
+        self.session_det = ort.InferenceSession(
+            path_or_bytes=onnx_det, providers=providers
+        )
+        self.session_pose = ort.InferenceSession(
+            path_or_bytes=onnx_pose, providers=providers
+        )
+
+    def __call__(self, oriImg):
+        det_result = inference_detector(self.session_det, oriImg)
+        keypoints, scores = inference_pose(self.session_pose, det_result, oriImg)
+
+        keypoints_info = np.concatenate((keypoints, scores[..., None]), axis=-1)
+        # compute neck joint
+        neck = np.mean(keypoints_info[:, [5, 6]], axis=1)
+        # neck score when visualizing pred
+        neck[:, 2:4] = np.logical_and(
+            keypoints_info[:, 5, 2:4] > 0.3, keypoints_info[:, 6, 2:4] > 0.3
+        ).astype(int)
+        new_keypoints_info = np.insert(keypoints_info, 17, neck, axis=1)
+        mmpose_idx = [17, 6, 8, 10, 7, 9, 12, 14, 16, 13, 15, 2, 1, 4, 3]
+        openpose_idx = [1, 2, 3, 4, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17]
+        new_keypoints_info[:, openpose_idx] = new_keypoints_info[:, mmpose_idx]
+        keypoints_info = new_keypoints_info
+
+        keypoints, scores = keypoints_info[..., :2], keypoints_info[..., 2]
+
+        return keypoints, scores

src/models/attention.py

@@ -0,0 +1,481 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
+
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.models.attention import AdaLayerNorm, Attention, FeedForward
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from einops import rearrange
+from torch import nn
+
+
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(
+                dim, max_seq_length=num_positional_embeddings
+            )
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(
+                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+            )
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(
+                    dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+                )
+            )
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim
+                if not double_self_attention
+                else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(
+                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+            )
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(
+                dim, cross_attention_dim, num_attention_heads, attention_head_dim
+            )
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = (
+            cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        )
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states
+            if self.only_cross_attention
+            else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = (
+                norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+            )
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class TemporalBasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        name=None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
+        self.unet_use_temporal_attention = unet_use_temporal_attention
+        self.name=name
+
+        # SC-Attn
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+        self.norm1 = (
+            AdaLayerNorm(dim, num_embeds_ada_norm)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(dim)
+        )
+
+        # Cross-Attn
+        if cross_attention_dim is not None:
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            self.attn2 = None
+
+        if cross_attention_dim is not None:
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+        else:
+            self.norm2 = None
+
+        # Feed-forward
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.norm3 = nn.LayerNorm(dim)
+        self.use_ada_layer_norm_zero = False
+        # Temp-Attn
+        assert unet_use_temporal_attention is not None
+        if unet_use_temporal_attention:
+            self.attn_temp = Attention(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+            nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
+            self.norm_temp = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        timestep=None,
+        attention_mask=None,
+        video_length=None,
+        self_attention_additional_feats=None,
+        mode=None,
+    ):
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep)
+            if self.use_ada_layer_norm
+            else self.norm1(hidden_states)
+        )
+        if self.name:
+            modify_norm_hidden_states = norm_hidden_states
+            if mode == "write":
+                self_attention_additional_feats[self.name]=norm_hidden_states
+            elif mode == "read" and self_attention_additional_feats:
+                ref_states = self_attention_additional_feats[self.name]
+                bank_fea = [
+                    rearrange(
+                        ref_states.unsqueeze(1).repeat(1, video_length, 1, 1),
+                        "b t l c -> (b t) l c",
+                    )
+                ]
+                modify_norm_hidden_states = torch.cat(
+                    [norm_hidden_states] + bank_fea, dim=1
+                )
+
+            if self.unet_use_cross_frame_attention:
+                hidden_states = (
+                    self.attn1(
+                        norm_hidden_states,
+                        attention_mask=attention_mask,
+                        encoder_hidden_states=modify_norm_hidden_states,
+                        video_length=video_length,
+                    )
+                    + hidden_states
+                )
+            else:
+                hidden_states = (
+                    self.attn1(
+                        norm_hidden_states, 
+                        encoder_hidden_states=modify_norm_hidden_states,
+                        attention_mask=attention_mask
+                    )
+                    + hidden_states
+                )
+        else:    
+            if self.unet_use_cross_frame_attention:
+                hidden_states = (
+                    self.attn1(
+                        norm_hidden_states,
+                        attention_mask=attention_mask,
+                        video_length=video_length,
+                    )
+                    + hidden_states
+                )
+            else:
+                hidden_states = (
+                    self.attn1(norm_hidden_states, attention_mask=attention_mask)
+                    + hidden_states
+                )
+
+        if self.attn2 is not None:
+            # Cross-Attention
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm2(hidden_states)
+            )
+            hidden_states = (
+                self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                )
+                + hidden_states
+            )
+
+        # Feed-forward
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+        # Temporal-Attention
+        if self.unet_use_temporal_attention:
+            d = hidden_states.shape[1]
+            hidden_states = rearrange(
+                hidden_states, "(b f) d c -> (b d) f c", f=video_length
+            )
+            norm_hidden_states = (
+                self.norm_temp(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm_temp(hidden_states)
+            )
+            hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

src/models/motion_module.py

@@ -0,0 +1,388 @@
+# Adapt from https://github.com/guoyww/AnimateDiff/blob/main/animatediff/models/motion_module.py
+import math
+from dataclasses import dataclass
+from typing import Callable, Optional
+
+import torch
+from diffusers.models.attention import FeedForward
+from diffusers.models.attention_processor import Attention, AttnProcessor
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from einops import rearrange, repeat
+from torch import nn
+
+
+def zero_module(module):
+    # Zero out the parameters of a module and return it.
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+@dataclass
+class TemporalTransformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+def get_motion_module(in_channels, motion_module_type: str, motion_module_kwargs: dict):
+    if motion_module_type == "Vanilla":
+        return VanillaTemporalModule(
+            in_channels=in_channels,
+            **motion_module_kwargs,
+        )
+    else:
+        raise ValueError
+
+
+class VanillaTemporalModule(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads=8,
+        num_transformer_block=2,
+        attention_block_types=("Temporal_Self", "Temporal_Self"),
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+        temporal_attention_dim_div=1,
+        zero_initialize=True,
+    ):
+        super().__init__()
+
+        self.temporal_transformer = TemporalTransformer3DModel(
+            in_channels=in_channels,
+            num_attention_heads=num_attention_heads,
+            attention_head_dim=in_channels
+            // num_attention_heads
+            // temporal_attention_dim_div,
+            num_layers=num_transformer_block,
+            attention_block_types=attention_block_types,
+            cross_frame_attention_mode=cross_frame_attention_mode,
+            temporal_position_encoding=temporal_position_encoding,
+            temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+        )
+
+        if zero_initialize:
+            self.temporal_transformer.proj_out = zero_module(
+                self.temporal_transformer.proj_out
+            )
+
+    def forward(
+        self,
+        input_tensor,
+        temb,
+        encoder_hidden_states,
+        attention_mask=None,
+        anchor_frame_idx=None,
+    ):
+        hidden_states = input_tensor
+        hidden_states = self.temporal_transformer(
+            hidden_states, encoder_hidden_states, attention_mask
+        )
+
+        output = hidden_states
+        return output
+
+
+class TemporalTransformer3DModel(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads,
+        attention_head_dim,
+        num_layers,
+        attention_block_types=(
+            "Temporal_Self",
+            "Temporal_Self",
+        ),
+        dropout=0.0,
+        norm_num_groups=32,
+        cross_attention_dim=768,
+        activation_fn="geglu",
+        attention_bias=False,
+        upcast_attention=False,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+    ):
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
+        )
+        self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TemporalTransformerBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    attention_block_types=attention_block_types,
+                    dropout=dropout,
+                    norm_num_groups=norm_num_groups,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    attention_bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    cross_frame_attention_mode=cross_frame_attention_mode,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+                for d in range(num_layers)
+            ]
+        )
+        self.proj_out = nn.Linear(inner_dim, in_channels)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        assert (
+            hidden_states.dim() == 5
+        ), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        inner_dim = hidden_states.shape[1]
+        hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+            batch, height * weight, inner_dim
+        )
+        hidden_states = self.proj_in(hidden_states)
+
+        # Transformer Blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                video_length=video_length,
+            )
+
+        # output
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = (
+            hidden_states.reshape(batch, height, weight, inner_dim)
+            .permute(0, 3, 1, 2)
+            .contiguous()
+        )
+
+        output = hidden_states + residual
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+
+        return output
+
+
+class TemporalTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_attention_heads,
+        attention_head_dim,
+        attention_block_types=(
+            "Temporal_Self",
+            "Temporal_Self",
+        ),
+        dropout=0.0,
+        norm_num_groups=32,
+        cross_attention_dim=768,
+        activation_fn="geglu",
+        attention_bias=False,
+        upcast_attention=False,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+    ):
+        super().__init__()
+
+        attention_blocks = []
+        norms = []
+
+        for block_name in attention_block_types:
+            attention_blocks.append(
+                VersatileAttention(
+                    attention_mode=block_name.split("_")[0],
+                    cross_attention_dim=cross_attention_dim
+                    if block_name.endswith("_Cross")
+                    else None,
+                    query_dim=dim,
+                    heads=num_attention_heads,
+                    dim_head=attention_head_dim,
+                    dropout=dropout,
+                    bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    cross_frame_attention_mode=cross_frame_attention_mode,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+            )
+            norms.append(nn.LayerNorm(dim))
+
+        self.attention_blocks = nn.ModuleList(attention_blocks)
+        self.norms = nn.ModuleList(norms)
+
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.ff_norm = nn.LayerNorm(dim)
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        video_length=None,
+    ):
+        for attention_block, norm in zip(self.attention_blocks, self.norms):
+            norm_hidden_states = norm(hidden_states)
+            hidden_states = (
+                attention_block(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states
+                    if attention_block.is_cross_attention
+                    else None,
+                    video_length=video_length,
+                )
+                + hidden_states
+            )
+
+        hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states
+
+        output = hidden_states
+        return output
+
+
+class PositionalEncoding(nn.Module):
+    def __init__(self, d_model, dropout=0.0, max_len=24):
+        super().__init__()
+        self.dropout = nn.Dropout(p=dropout)
+        position = torch.arange(max_len).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model)
+        )
+        pe = torch.zeros(1, max_len, d_model)
+        pe[0, :, 0::2] = torch.sin(position * div_term)
+        pe[0, :, 1::2] = torch.cos(position * div_term)
+        self.register_buffer("pe", pe)
+
+    def forward(self, x):
+        x = x + self.pe[:, : x.size(1)]
+        return self.dropout(x)
+
+
+class VersatileAttention(Attention):
+    def __init__(
+        self,
+        attention_mode=None,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+        *args,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        assert attention_mode == "Temporal"
+
+        self.attention_mode = attention_mode
+        self.is_cross_attention = kwargs["cross_attention_dim"] is not None
+
+        self.pos_encoder = (
+            PositionalEncoding(
+                kwargs["query_dim"],
+                dropout=0.0,
+                max_len=temporal_position_encoding_max_len,
+            )
+            if (temporal_position_encoding and attention_mode == "Temporal")
+            else None
+        )
+
+    def extra_repr(self):
+        return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}"
+
+    def set_use_memory_efficient_attention_xformers(
+        self,
+        use_memory_efficient_attention_xformers: bool,
+        attention_op: Optional[Callable] = None,
+    ):
+        if use_memory_efficient_attention_xformers:
+            if not is_xformers_available():
+                raise ModuleNotFoundError(
+                    (
+                        "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+                        " xformers"
+                    ),
+                    name="xformers",
+                )
+            elif not torch.cuda.is_available():
+                raise ValueError(
+                    "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
+                    " only available for GPU "
+                )
+            else:
+                try:
+                    # Make sure we can run the memory efficient attention
+                    _ = xformers.ops.memory_efficient_attention(
+                        torch.randn((1, 2, 40), device="cuda"),
+                        torch.randn((1, 2, 40), device="cuda"),
+                        torch.randn((1, 2, 40), device="cuda"),
+                    )
+                except Exception as e:
+                    raise e
+
+            # XFormersAttnProcessor corrupts video generation and work with Pytorch 1.13.
+            # Pytorch 2.0.1 AttnProcessor works the same as XFormersAttnProcessor in Pytorch 1.13.
+            # You don't need XFormersAttnProcessor here.
+            # processor = XFormersAttnProcessor(
+            #     attention_op=attention_op,
+            # )
+            processor = AttnProcessor()
+        else:
+            processor = AttnProcessor()
+
+        self.set_processor(processor)
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        video_length=None,
+        **cross_attention_kwargs,
+    ):
+        if self.attention_mode == "Temporal":
+            d = hidden_states.shape[1]  # d means HxW
+            hidden_states = rearrange(
+                hidden_states, "(b f) d c -> (b d) f c", f=video_length
+            )
+
+            if self.pos_encoder is not None:
+                hidden_states = self.pos_encoder(hidden_states)
+
+            encoder_hidden_states = (
+                repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d)
+                if encoder_hidden_states is not None
+                else encoder_hidden_states
+            )
+
+        else:
+            raise NotImplementedError
+
+        hidden_states = self.processor(
+            self,
+            hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+        if self.attention_mode == "Temporal":
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

src/models/mutual_self_attention.py

@@ -0,0 +1,365 @@
+# Adapted from https://github.com/magic-research/magic-animate/blob/main/magicanimate/models/mutual_self_attention.py
+from typing import Any, Dict, Optional
+
+import torch
+from einops import rearrange
+
+from src.models.attention import TemporalBasicTransformerBlock
+
+#from .attention import BasicTransformerBlock
+from diffusers.models.attention import BasicTransformerBlock
+
+def torch_dfs(model: torch.nn.Module):
+    result = [model]
+    for child in model.children():
+        result += torch_dfs(child)
+    return result
+
+
+class ReferenceAttentionControl:
+    def __init__(
+        self,
+        unet,
+        mode="write",
+        do_classifier_free_guidance=False,
+        attention_auto_machine_weight=float("inf"),
+        gn_auto_machine_weight=1.0,
+        style_fidelity=1.0,
+        reference_attn=True,
+        reference_adain=False,
+        fusion_blocks="midup",
+        batch_size=1,
+    ) -> None:
+        # 10. Modify self attention and group norm
+        self.unet = unet
+        assert mode in ["read", "write"]
+        assert fusion_blocks in ["midup", "full"]
+        self.reference_attn = reference_attn
+        self.reference_adain = reference_adain
+        self.fusion_blocks = fusion_blocks
+        self.register_reference_hooks(
+            mode,
+            do_classifier_free_guidance,
+            attention_auto_machine_weight,
+            gn_auto_machine_weight,
+            style_fidelity,
+            reference_attn,
+            reference_adain,
+            fusion_blocks,
+            batch_size=batch_size,
+        )
+
+    def register_reference_hooks(
+        self,
+        mode,
+        do_classifier_free_guidance,
+        attention_auto_machine_weight,
+        gn_auto_machine_weight,
+        style_fidelity,
+        reference_attn,
+        reference_adain,
+        dtype=torch.float16,
+        batch_size=1,
+        num_images_per_prompt=1,
+        device=torch.device("cpu"),
+        fusion_blocks="midup",
+    ):
+        MODE = mode
+        do_classifier_free_guidance = do_classifier_free_guidance
+        attention_auto_machine_weight = attention_auto_machine_weight
+        gn_auto_machine_weight = gn_auto_machine_weight
+        style_fidelity = style_fidelity
+        reference_attn = reference_attn
+        reference_adain = reference_adain
+        fusion_blocks = fusion_blocks
+        num_images_per_prompt = num_images_per_prompt
+        dtype = dtype
+        if do_classifier_free_guidance:
+            uc_mask = (
+                torch.Tensor(
+                    [1] * batch_size * num_images_per_prompt * 16
+                    + [0] * batch_size * num_images_per_prompt * 16
+                )
+                .to(device)
+                .bool()
+            )
+        else:
+            uc_mask = (
+                torch.Tensor([0] * batch_size * num_images_per_prompt * 2)
+                .to(device)
+                .bool()
+            )
+
+        def hacked_basic_transformer_inner_forward(
+            self,
+            hidden_states: torch.FloatTensor,
+            attention_mask: Optional[torch.FloatTensor] = None,
+            encoder_hidden_states: Optional[torch.FloatTensor] = None,
+            encoder_attention_mask: Optional[torch.FloatTensor] = None,
+            timestep: Optional[torch.LongTensor] = None,
+            cross_attention_kwargs: Dict[str, Any] = None,
+            class_labels: Optional[torch.LongTensor] = None,
+            video_length=None,
+            self_attention_additional_feats=None,
+            mode=None,
+        ):
+            if self.use_ada_layer_norm:  # False
+                norm_hidden_states = self.norm1(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero:
+                (
+                    norm_hidden_states,
+                    gate_msa,
+                    shift_mlp,
+                    scale_mlp,
+                    gate_mlp,
+                ) = self.norm1(
+                    hidden_states,
+                    timestep,
+                    class_labels,
+                    hidden_dtype=hidden_states.dtype,
+                )
+            else:
+                norm_hidden_states = self.norm1(hidden_states)
+
+            # 1. Self-Attention
+            # self.only_cross_attention = False
+            cross_attention_kwargs = (
+                cross_attention_kwargs if cross_attention_kwargs is not None else {}
+            )
+            if self.only_cross_attention:
+                attn_output = self.attn1(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states
+                    if self.only_cross_attention
+                    else None,
+                    attention_mask=attention_mask,
+                    **cross_attention_kwargs,
+                )
+            else:
+                if MODE == "write":
+                    self.bank.append(norm_hidden_states.clone())
+                    attn_output = self.attn1(
+                        norm_hidden_states,
+                        encoder_hidden_states=encoder_hidden_states
+                        if self.only_cross_attention
+                        else None,
+                        attention_mask=attention_mask,
+                        **cross_attention_kwargs,
+                    )
+                if MODE == "read":
+                    bank_fea = [
+                        rearrange(
+                            d.unsqueeze(1).repeat(1, video_length, 1, 1),
+                            "b t l c -> (b t) l c",
+                        )
+                        for d in self.bank
+                    ]
+                    modify_norm_hidden_states = torch.cat(
+                        [norm_hidden_states] + bank_fea, dim=1
+                    )
+                    hidden_states_uc = (
+                        self.attn1(
+                            norm_hidden_states,
+                            encoder_hidden_states=modify_norm_hidden_states,
+                            attention_mask=attention_mask,
+                        )
+                        + hidden_states
+                    )
+                    if do_classifier_free_guidance:
+                        hidden_states_c = hidden_states_uc.clone()
+                        _uc_mask = uc_mask.clone()
+                        if hidden_states.shape[0] != _uc_mask.shape[0]:
+                            _uc_mask = (
+                                torch.Tensor(
+                                    [1] * (hidden_states.shape[0] // 2)
+                                    + [0] * (hidden_states.shape[0] // 2)
+                                )
+                                .to(device)
+                                .bool()
+                            )
+                        hidden_states_c[_uc_mask] = (
+                            self.attn1(
+                                norm_hidden_states[_uc_mask],
+                                encoder_hidden_states=norm_hidden_states[_uc_mask],
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states[_uc_mask]
+                        )
+                        hidden_states = hidden_states_c.clone()
+                    else:
+                        hidden_states = hidden_states_uc
+
+                    # self.bank.clear()
+                    if self.attn2 is not None:
+                        # Cross-Attention
+                        norm_hidden_states = (
+                            self.norm2(hidden_states, timestep)
+                            if self.use_ada_layer_norm
+                            else self.norm2(hidden_states)
+                        )
+                        hidden_states = (
+                            self.attn2(
+                                norm_hidden_states,
+                                encoder_hidden_states=encoder_hidden_states,
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states
+                        )
+
+                    # Feed-forward
+                    hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+                    # Temporal-Attention
+                    if self.unet_use_temporal_attention:
+                        d = hidden_states.shape[1]
+                        hidden_states = rearrange(
+                            hidden_states, "(b f) d c -> (b d) f c", f=video_length
+                        )
+                        norm_hidden_states = (
+                            self.norm_temp(hidden_states, timestep)
+                            if self.use_ada_layer_norm
+                            else self.norm_temp(hidden_states)
+                        )
+                        hidden_states = (
+                            self.attn_temp(norm_hidden_states) + hidden_states
+                        )
+                        hidden_states = rearrange(
+                            hidden_states, "(b d) f c -> (b f) d c", d=d
+                        )
+
+                    return hidden_states
+
+            if self.use_ada_layer_norm_zero:
+                attn_output = gate_msa.unsqueeze(1) * attn_output
+            hidden_states = attn_output + hidden_states
+
+            if self.attn2 is not None:
+                norm_hidden_states = (
+                    self.norm2(hidden_states, timestep)
+                    if self.use_ada_layer_norm
+                    else self.norm2(hidden_states)
+                )
+
+                # 2. Cross-Attention
+                attn_output = self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=encoder_attention_mask,
+                    **cross_attention_kwargs,
+                )
+                hidden_states = attn_output + hidden_states
+
+            # 3. Feed-forward
+            norm_hidden_states = self.norm3(hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                norm_hidden_states = (
+                    norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+                )
+
+            ff_output = self.ff(norm_hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+            hidden_states = ff_output + hidden_states
+
+            return hidden_states
+
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            attn_modules = sorted(
+                attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+
+            for i, module in enumerate(attn_modules):
+                module._original_inner_forward = module.forward
+                if isinstance(module, BasicTransformerBlock):
+                    module.forward = hacked_basic_transformer_inner_forward.__get__(
+                        module, BasicTransformerBlock
+                    )
+                if isinstance(module, TemporalBasicTransformerBlock):
+                    module.forward = hacked_basic_transformer_inner_forward.__get__(
+                        module, TemporalBasicTransformerBlock
+                    )
+
+                module.bank = []
+                module.attn_weight = float(i) / float(len(attn_modules))
+
+    def update(self, writer, dtype=torch.float16):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, TemporalBasicTransformerBlock)
+                ]
+                writer_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(writer.unet.mid_block)
+                        + torch_dfs(writer.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, TemporalBasicTransformerBlock)
+                ]
+                writer_attn_modules = [
+                    module
+                    for module in torch_dfs(writer.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            reader_attn_modules = sorted(
+                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            writer_attn_modules = sorted(
+                writer_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            for r, w in zip(reader_attn_modules, writer_attn_modules):
+                r.bank = [v.clone().to(dtype) for v in w.bank]
+                # w.bank.clear()
+
+    def clear(self):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            reader_attn_modules = sorted(
+                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            for r in reader_attn_modules:
+                r.bank.clear()

src/models/pose_guider.py

@@ -0,0 +1,57 @@
+from typing import Tuple
+
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.nn.init as init
+from diffusers.models.modeling_utils import ModelMixin
+
+from src.models.motion_module import zero_module
+from src.models.resnet import InflatedConv3d
+
+
+class PoseGuider(ModelMixin):
+    def __init__(
+        self,
+        conditioning_embedding_channels: int,
+        conditioning_channels: int = 3,
+        block_out_channels: Tuple[int] = (16, 32, 64, 128),
+    ):
+        super().__init__()
+        self.conv_in = InflatedConv3d(
+            conditioning_channels, block_out_channels[0], kernel_size=3, padding=1
+        )
+
+        self.blocks = nn.ModuleList([])
+
+        for i in range(len(block_out_channels) - 1):
+            channel_in = block_out_channels[i]
+            channel_out = block_out_channels[i + 1]
+            self.blocks.append(
+                InflatedConv3d(channel_in, channel_in, kernel_size=3, padding=1)
+            )
+            self.blocks.append(
+                InflatedConv3d(
+                    channel_in, channel_out, kernel_size=3, padding=1, stride=2
+                )
+            )
+
+        self.conv_out = zero_module(
+            InflatedConv3d(
+                block_out_channels[-1],
+                conditioning_embedding_channels,
+                kernel_size=3,
+                padding=1,
+            )
+        )
+
+    def forward(self, conditioning):
+        embedding = self.conv_in(conditioning)
+        embedding = F.silu(embedding)
+
+        for block in self.blocks:
+            embedding = block(embedding)
+            embedding = F.silu(embedding)
+
+        embedding = self.conv_out(embedding)
+
+        return embedding

src/models/resnet.py

@@ -0,0 +1,252 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/resnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+
+class InflatedConv3d(nn.Conv2d):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class InflatedGroupNorm(nn.GroupNorm):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class Upsample3D(nn.Module):
+    def __init__(
+        self,
+        channels,
+        use_conv=False,
+        use_conv_transpose=False,
+        out_channels=None,
+        name="conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        conv = None
+        if use_conv_transpose:
+            raise NotImplementedError
+        elif use_conv:
+            self.conv = InflatedConv3d(self.channels, self.out_channels, 3, padding=1)
+
+    def forward(self, hidden_states, output_size=None):
+        assert hidden_states.shape[1] == self.channels
+
+        if self.use_conv_transpose:
+            raise NotImplementedError
+
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
+        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+        if hidden_states.shape[0] >= 64:
+            hidden_states = hidden_states.contiguous()
+
+        # if `output_size` is passed we force the interpolation output
+        # size and do not make use of `scale_factor=2`
+        if output_size is None:
+            hidden_states = F.interpolate(
+                hidden_states, scale_factor=[1.0, 2.0, 2.0], mode="nearest"
+            )
+        else:
+            hidden_states = F.interpolate(
+                hidden_states, size=output_size, mode="nearest"
+            )
+
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
+        # if self.use_conv:
+        #     if self.name == "conv":
+        #         hidden_states = self.conv(hidden_states)
+        #     else:
+        #         hidden_states = self.Conv2d_0(hidden_states)
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class Downsample3D(nn.Module):
+    def __init__(
+        self, channels, use_conv=False, out_channels=None, padding=1, name="conv"
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+
+        if use_conv:
+            self.conv = InflatedConv3d(
+                self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            raise NotImplementedError
+
+    def forward(self, hidden_states):
+        assert hidden_states.shape[1] == self.channels
+        if self.use_conv and self.padding == 0:
+            raise NotImplementedError
+
+        assert hidden_states.shape[1] == self.channels
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class ResnetBlock3D(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout=0.0,
+        temb_channels=512,
+        groups=32,
+        groups_out=None,
+        pre_norm=True,
+        eps=1e-6,
+        non_linearity="swish",
+        time_embedding_norm="default",
+        output_scale_factor=1.0,
+        use_in_shortcut=None,
+        use_inflated_groupnorm=None,
+    ):
+        super().__init__()
+        self.pre_norm = pre_norm
+        self.pre_norm = True
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.time_embedding_norm = time_embedding_norm
+        self.output_scale_factor = output_scale_factor
+
+        if groups_out is None:
+            groups_out = groups
+
+        assert use_inflated_groupnorm != None
+        if use_inflated_groupnorm:
+            self.norm1 = InflatedGroupNorm(
+                num_groups=groups, num_channels=in_channels, eps=eps, affine=True
+            )
+        else:
+            self.norm1 = torch.nn.GroupNorm(
+                num_groups=groups, num_channels=in_channels, eps=eps, affine=True
+            )
+
+        self.conv1 = InflatedConv3d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+        if temb_channels is not None:
+            if self.time_embedding_norm == "default":
+                time_emb_proj_out_channels = out_channels
+            elif self.time_embedding_norm == "scale_shift":
+                time_emb_proj_out_channels = out_channels * 2
+            else:
+                raise ValueError(
+                    f"unknown time_embedding_norm : {self.time_embedding_norm} "
+                )
+
+            self.time_emb_proj = torch.nn.Linear(
+                temb_channels, time_emb_proj_out_channels
+            )
+        else:
+            self.time_emb_proj = None
+
+        if use_inflated_groupnorm:
+            self.norm2 = InflatedGroupNorm(
+                num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
+            )
+        else:
+            self.norm2 = torch.nn.GroupNorm(
+                num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
+            )
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = InflatedConv3d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+        if non_linearity == "swish":
+            self.nonlinearity = lambda x: F.silu(x)
+        elif non_linearity == "mish":
+            self.nonlinearity = Mish()
+        elif non_linearity == "silu":
+            self.nonlinearity = nn.SiLU()
+
+        self.use_in_shortcut = (
+            self.in_channels != self.out_channels
+            if use_in_shortcut is None
+            else use_in_shortcut
+        )
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = InflatedConv3d(
+                in_channels, out_channels, kernel_size=1, stride=1, padding=0
+            )
+
+    def forward(self, input_tensor, temb):
+        hidden_states = input_tensor
+
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv1(hidden_states)
+
+        if temb is not None:
+            temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if temb is not None and self.time_embedding_norm == "default":
+            hidden_states = hidden_states + temb
+
+        hidden_states = self.norm2(hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+class Mish(torch.nn.Module):
+    def forward(self, hidden_states):
+        return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))

src/models/transformer_2d.py

@@ -0,0 +1,396 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformer_2d.py
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.embeddings import CaptionProjection
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
+from torch import nn
+
+from .attention import BasicTransformerBlock
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+    ref_feature: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate(
+                "norm_type!=num_embeds_ada_norm",
+                "1.0.0",
+                deprecation_message,
+                standard_warn=False,
+            )
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif (
+            not self.is_input_continuous
+            and not self.is_input_vectorized
+            and not self.is_input_patches
+        ):
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups,
+            num_channels=in_channels,
+            eps=1e-6,
+            affine=True,
+        )
+        if use_linear_projection:
+            self.proj_in = linear_cls(in_channels, inner_dim)
+        else:
+            self.proj_in = conv_cls(
+                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        # TODO: should use out_channels for continuous projections
+        if use_linear_projection:
+            self.proj_out = linear_cls(inner_dim, in_channels)
+        else:
+            self.proj_out = conv_cls(
+                inner_dim, in_channels, kernel_size=1, stride=1, padding=0
+            )
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(
+                inner_dim, use_additional_conditions=self.use_additional_conditions
+            )
+
+        self.caption_projection = None
+        if caption_channels is not None:
+            self.caption_projection = CaptionProjection(
+                in_features=caption_channels, hidden_size=inner_dim
+            )
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (
+                1 - encoder_attention_mask.to(hidden_states.dtype)
+            ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # Retrieve lora scale.
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        # 1. Input
+        batch, _, height, width = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale)
+                if not USE_PEFT_BACKEND
+                else self.proj_in(hidden_states)
+            )
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * width, inner_dim
+            )
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * width, inner_dim
+            )
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale)
+                if not USE_PEFT_BACKEND
+                else self.proj_in(hidden_states)
+            )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(
+                batch_size, -1, hidden_states.shape[-1]
+            )
+
+        ref_feature = hidden_states.reshape(batch, height, width, inner_dim)
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            if not self.use_linear_projection:
+                hidden_states = (
+                    hidden_states.reshape(batch, height, width, inner_dim)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+            else:
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+                hidden_states = (
+                    hidden_states.reshape(batch, height, width, inner_dim)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+
+            output = hidden_states + residual
+        if not return_dict:
+            return (output, ref_feature)
+
+        return Transformer2DModelOutput(sample=output, ref_feature=ref_feature)

src/models/transformer_3d.py

@@ -0,0 +1,202 @@
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models import ModelMixin
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from einops import rearrange, repeat
+from torch import nn
+
+from .attention import TemporalBasicTransformerBlock
+
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        name=None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # Define input layers
+        self.in_channels = in_channels
+        self.name=name
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
+        )
+        if use_linear_projection:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_in = nn.Conv2d(
+                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+            )
+
+        # Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TemporalBasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                    name=f"{self.name}_{d}_TransformerBlock" if self.name else None,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        if use_linear_projection:
+            self.proj_out = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_out = nn.Conv2d(
+                inner_dim, in_channels, kernel_size=1, stride=1, padding=0
+            )
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        self_attention_additional_feats=None,
+        mode=None,
+        timestep=None,
+        return_dict: bool = True,
+    ):
+        # Input
+        assert (
+            hidden_states.dim() == 5
+        ), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+        if encoder_hidden_states.shape[0] != hidden_states.shape[0]:
+            encoder_hidden_states = repeat(
+                encoder_hidden_states, "b n c -> (b f) n c", f=video_length
+            )
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = self.proj_in(hidden_states)
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * weight, inner_dim
+            )
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * weight, inner_dim
+            )
+            hidden_states = self.proj_in(hidden_states)
+
+        # Blocks
+        for i, block in enumerate(self.transformer_blocks):
+            
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                # if hasattr(self.block, 'bank') and len(self.block.bank) > 0:
+                #     hidden_states
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    timestep=timestep,
+                    attention_mask=None,
+                    video_length=video_length,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=mode,
+                )
+            else:
+
+                hidden_states = block(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    timestep=timestep,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=mode,
+                    video_length=video_length,
+                )
+
+        # Output
+        if not self.use_linear_projection:
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim)
+                .permute(0, 3, 1, 2)
+                .contiguous()
+            )
+            hidden_states = self.proj_out(hidden_states)
+        else:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim)
+                .permute(0, 3, 1, 2)
+                .contiguous()
+            )
+
+        output = hidden_states + residual
+
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+        if not return_dict:
+            return (output,)
+
+        return Transformer3DModelOutput(sample=output)

src/models/unet_2d_blocks.py

@@ -0,0 +1,1074 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+from typing import Any, Dict, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import Attention
+from diffusers.models.dual_transformer_2d import DualTransformer2DModel
+from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
+from diffusers.utils import is_torch_version, logging
+from diffusers.utils.torch_utils import apply_freeu
+from torch import nn
+
+from .transformer_2d import Transformer2DModel
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def get_down_block(
+    down_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    temb_channels: int,
+    add_downsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    downsample_padding: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    downsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+):
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    down_block_type = (
+        down_block_type[7:]
+        if down_block_type.startswith("UNetRes")
+        else down_block_type
+    )
+    if down_block_type == "DownBlock2D":
+        return DownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "CrossAttnDownBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnDownBlock2D"
+            )
+        return CrossAttnDownBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    prev_output_channel: int,
+    temb_channels: int,
+    add_upsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    resolution_idx: Optional[int] = None,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    upsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+) -> nn.Module:
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    up_block_type = (
+        up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    )
+    if up_block_type == "UpBlock2D":
+        return UpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif up_block_type == "CrossAttnUpBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnUpBlock2D"
+            )
+        return CrossAttnUpBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class AutoencoderTinyBlock(nn.Module):
+    """
+    Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
+    blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        out_channels (`int`): The number of output channels.
+        act_fn (`str`):
+            ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
+
+    Returns:
+        `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
+        `out_channels`.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, act_fn: str):
+        super().__init__()
+        act_fn = get_activation(act_fn)
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+        )
+        self.skip = (
+            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+        self.fuse = nn.ReLU()
+
+    def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        return self.fuse(self.conv(x) + self.skip(x))
+
+
+class UNetMidBlock2D(nn.Module):
+    """
+    A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        temb_channels (`int`): The number of temporal embedding channels.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
+        num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
+        resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
+        resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
+            The type of normalization to apply to the time embeddings. This can help to improve the performance of the
+            model on tasks with long-range temporal dependencies.
+        resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
+        resnet_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use in the group normalization layers of the resnet blocks.
+        attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
+        resnet_pre_norm (`bool`, *optional*, defaults to `True`):
+            Whether to use pre-normalization for the resnet blocks.
+        add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
+        attention_head_dim (`int`, *optional*, defaults to 1):
+            Dimension of a single attention head. The number of attention heads is determined based on this value and
+            the number of input channels.
+        output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
+
+    Returns:
+        `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
+        in_channels, height, width)`.
+
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",  # default, spatial
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        attn_groups: Optional[int] = None,
+        resnet_pre_norm: bool = True,
+        add_attention: bool = True,
+        attention_head_dim: int = 1,
+        output_scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+        self.add_attention = add_attention
+
+        if attn_groups is None:
+            attn_groups = (
+                resnet_groups if resnet_time_scale_shift == "default" else None
+            )
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        if attention_head_dim is None:
+            logger.warn(
+                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+            )
+            attention_head_dim = in_channels
+
+        for _ in range(num_layers):
+            if self.add_attention:
+                attentions.append(
+                    Attention(
+                        in_channels,
+                        heads=in_channels // attention_head_dim,
+                        dim_head=attention_head_dim,
+                        rescale_output_factor=output_scale_factor,
+                        eps=resnet_eps,
+                        norm_num_groups=attn_groups,
+                        spatial_norm_dim=temb_channels
+                        if resnet_time_scale_shift == "spatial"
+                        else None,
+                        residual_connection=True,
+                        bias=True,
+                        upcast_softmax=True,
+                        _from_deprecated_attn_block=True,
+                    )
+                )
+            else:
+                attentions.append(None)
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(
+        self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None
+    ) -> torch.FloatTensor:
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                hidden_states = attn(hidden_states, temb=temb)
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        output_scale_factor: float = 1.0,
+        cross_attention_dim: int = 1280,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+
+        # support for variable transformer layers per block
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for i in range(num_layers):
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+        return hidden_states
+
+
+class CrossAttnDownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        downsample_padding: int = 1,
+        add_downsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        additional_residuals: Optional[torch.FloatTensor] = None,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        blocks = list(zip(self.resnets, self.attentions))
+
+        for i, (resnet, attn) in enumerate(blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+
+            # apply additional residuals to the output of the last pair of resnet and attention blocks
+            if i == len(blocks) - 1 and additional_residuals is not None:
+                hidden_states = hidden_states + additional_residuals
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_downsample: bool = True,
+        downsample_padding: int = 1,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        scale: float = 1.0,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        upsample_size: Optional[int] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(
+                    hidden_states, upsample_size, scale=lora_scale
+                )
+
+        return hidden_states
+
+
+class UpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        upsample_size: Optional[int] = None,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        for resnet in self.resnets:
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
+
+        return hidden_states

src/models/unet_2d_condition.py

@@ -0,0 +1,1308 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    PositionNet,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    deprecate,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+
+from .unet_2d_blocks import (
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+    ref_features: Tuple[torch.FloatTensor] = None
+
+
+class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+       reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+        *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+        *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = (
+            "UpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: int = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads=64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(
+            only_cross_attention
+        ) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if (
+            isinstance(transformer_layers_per_block, list)
+            and reverse_transformer_layers_per_block is None
+        ):
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError(
+                        "Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet."
+                    )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=conv_in_kernel,
+            padding=conv_in_padding,
+        )
+
+        # time
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(
+                    f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}."
+                )
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2,
+                set_W_to_weight=False,
+                log=False,
+                flip_sin_to_cos=flip_sin_to_cos,
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(
+                block_out_channels[0], flip_sin_to_cos, freq_shift
+            )
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info(
+                "encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined."
+            )
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(
+                timestep_input_dim, time_embed_dim, act_fn=act_fn
+            )
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim,
+                time_embed_dim,
+                num_heads=addition_embed_type_num_heads,
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim,
+                image_embed_dim=cross_attention_dim,
+                time_embed_dim=time_embed_dim,
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(
+                addition_time_embed_dim, flip_sin_to_cos, freq_shift
+            )
+            self.add_embedding = TimestepEmbedding(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(
+                image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(
+                image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type is not None:
+            raise ValueError(
+                f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'."
+            )
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(
+                down_block_types
+            )
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i]
+                if attention_head_dim[i] is not None
+                else output_channel,
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                attention_type=attention_type,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            raise NotImplementedError(f"Unsupport mid_block_type: {mid_block_type}")
+        elif mid_block_type == "UNetMidBlock2D":
+            self.mid_block = UNetMidBlock2D(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                num_layers=0,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                add_attention=False,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[
+                min(i + 1, len(block_out_channels) - 1)
+            ]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i]
+                if attention_head_dim[i] is not None
+                else output_channel,
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+        self.conv_norm_out = None
+
+        conv_out_padding = (conv_out_kernel - 1) // 2
+        # self.conv_out = nn.Conv2d(
+        #     block_out_channels[0],
+        #     out_channels,
+        #     kernel_size=conv_out_kernel,
+        #     padding=conv_out_padding,
+        # )
+
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(
+                cross_attention_dim, list
+            ):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = PositionNet(
+                positive_len=positive_len,
+                out_dim=cross_attention_dim,
+                feature_type=feature_type,
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(
+                    return_deprecated_lora=True
+                )
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(
+        self,
+        processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]],
+        _remove_lora=False,
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(
+                        processor.pop(f"{name}.processor"), _remove_lora=_remove_lora
+                    )
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(
+            proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS
+            for proc in self.attn_processors.values()
+        ):
+            processor = AttnAddedKVProcessor()
+        elif all(
+            proc.__class__ in CROSS_ATTENTION_PROCESSORS
+            for proc in self.attn_processors.values()
+        ):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = (
+            num_sliceable_layers * [slice_size]
+            if not isinstance(slice_size, list)
+            else slice_size
+        )
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(
+            module: torch.nn.Module, slice_size: List[int]
+        ):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1, s2, b1, b2):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if (
+                    hasattr(upsample_block, k)
+                    or getattr(upsample_block, k, None) is not None
+                ):
+                    setattr(upsample_block, k, None)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+                example from ControlNet side model(s)
+            mid_block_additional_residual (`torch.Tensor`, *optional*):
+                additional residual to be added to UNet mid block output, for example from ControlNet side model
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (
+                1 - encoder_attention_mask.to(sample.dtype)
+            ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError(
+                    "class_labels should be provided when num_class_embeds > 0"
+                )
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if (
+                "image_embeds" not in added_cond_kwargs
+                or "hint" not in added_cond_kwargs
+            ):
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb, hint = self.add_embedding(image_embs, hint)
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "text_proj"
+        ):
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "text_image_proj"
+        ):
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(
+                encoder_hidden_states, image_embeds
+            )
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "image_proj"
+        ):
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "ip_image_proj"
+        ):
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds).to(
+                encoder_hidden_states.dtype
+            )
+            encoder_hidden_states = torch.cat(
+                [encoder_hidden_states, image_embeds], dim=1
+            )
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if (
+            cross_attention_kwargs is not None
+            and cross_attention_kwargs.get("gligen", None) is not None
+        ):
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {
+                "objs": self.position_net(**gligen_args)
+            }
+
+        # 3. down
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = (
+            mid_block_additional_residual is not None
+            and down_block_additional_residuals is not None
+        )
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if (
+            not is_adapter
+            and mid_block_additional_residual is None
+            and down_block_additional_residuals is not None
+        ):
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        tot_referece_features = ()
+        for downsample_block in self.down_blocks:
+            if (
+                hasattr(downsample_block, "has_cross_attention")
+                and downsample_block.has_cross_attention
+            ):
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals[
+                        "additional_residuals"
+                    ] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample, temb=emb, scale=lora_scale
+                )
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = (
+                    down_block_res_sample + down_block_additional_residual
+                )
+                new_down_block_res_samples = new_down_block_res_samples + (
+                    down_block_res_sample,
+                )
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if (
+                hasattr(self.mid_block, "has_cross_attention")
+                and self.mid_block.has_cross_attention
+            ):
+                sample = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[
+                : -len(upsample_block.resnets)
+            ]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if (
+                hasattr(upsample_block, "has_cross_attention")
+                and upsample_block.has_cross_attention
+            ):
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    scale=lora_scale,
+                )
+
+        # 6. post-process
+        # if self.conv_norm_out:
+        #     sample = self.conv_norm_out(sample)
+        #     sample = self.conv_act(sample)
+        # sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

src/models/unet_3d.py

@@ -0,0 +1,707 @@
+# Adapted from https://github.com/guoyww/AnimateDiff/blob/main/animatediff/models/unet_blocks.py
+
+from collections import OrderedDict
+from dataclasses import dataclass
+from os import PathLike
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.attention_processor import AttentionProcessor
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+try:
+    from diffusers.modeling_utils import ModelMixin
+except:
+    from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import SAFETENSORS_WEIGHTS_NAME, WEIGHTS_NAME, BaseOutput, logging
+from safetensors.torch import load_file
+
+from .resnet import InflatedConv3d, InflatedGroupNorm
+from .unet_3d_blocks import UNetMidBlock3DCrossAttn, get_down_block, get_up_block
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+from einops import rearrange
+from scipy.io import loadmat, savemat
+
+
+@dataclass
+class UNet3DConditionOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+class UNet3DConditionModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ),
+        mid_block_type: str = "UNetMidBlock3DCrossAttn",
+        up_block_types: Tuple[str] = (
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: int = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: int = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: int = 1280,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        use_inflated_groupnorm=False,
+        # Additional
+        use_motion_module=False,  ######
+        motion_module_resolutions=(1, 2, 4, 8),  ####
+        motion_module_mid_block=False,   #####
+        motion_module_decoder_only=False,   #####
+        motion_module_type=None,  #####
+        motion_module_kwargs={},   #####
+        unet_use_cross_frame_attention=None,  #####
+        unet_use_temporal_attention=None,  #####
+        mode=None,  #####
+        task_type="action",   #####
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        time_embed_dim = block_out_channels[0] * 4
+
+        # input
+        self.conv_in = InflatedConv3d(
+            in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1)
+        )
+
+        # time
+        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+        timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        self.down_blocks = nn.ModuleList([])
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            if task_type == "action":
+                name_index, mid_name = None, None
+            else:
+                name_index, mid_name = i, "MidBlock"
+            res = 2**i
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module
+                and (res in motion_module_resolutions)
+                and (not motion_module_decoder_only),
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+                name_index=name_index,  #####
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        
+        if mid_block_type == "UNetMidBlock3DCrossAttn":
+            self.mid_block = UNetMidBlock3DCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module and motion_module_mid_block,
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+                name=mid_name,
+            )
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the videos
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_attention_head_dim = list(reversed(attention_head_dim))
+        only_cross_attention = list(reversed(only_cross_attention))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            res = 2 ** (3 - i)
+            is_final_block = i == len(block_out_channels) - 1
+            
+            if task_type == "action":
+                name_index = None
+            else:
+                name_index = i
+            
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[
+                min(i + 1, len(block_out_channels) - 1)
+            ]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=layers_per_block + 1,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=reversed_attention_head_dim[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module
+                and (res in motion_module_resolutions),
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+                name_index=name_index,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if use_inflated_groupnorm:
+            self.conv_norm_out = InflatedGroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+        else:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+        self.conv_act = nn.SiLU()
+        self.conv_out = InflatedConv3d(
+            block_out_channels[0], out_channels, kernel_size=3, padding=1
+        )
+        
+        self.mode = mode
+
+    @property
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "set_processor"):
+                processors[f"{name}.processor"] = module.processor
+
+            for sub_name, child in module.named_children():
+                if "temporal_transformer" not in sub_name:
+                    fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            if "temporal_transformer" not in name:
+                fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_slicable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_slicable_dims(module)
+
+        num_slicable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_slicable_layers * [1]
+
+        slice_size = (
+            num_slicable_layers * [slice_size]
+            if not isinstance(slice_size, list)
+            else slice_size
+        )
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(
+            module: torch.nn.Module, slice_size: List[int]
+        ):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                if "temporal_transformer" not in sub_name:
+                    fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            if "temporal_transformer" not in name:
+                fn_recursive_attn_processor(name, module, processor)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        pose_cond_fea: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        self_attention_additional_feats = None,
+    ) -> Union[UNet3DConditionOutput, Tuple]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
+            timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
+            encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=self.dtype)
+        emb = self.time_embedding(t_emb)
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError(
+                    "class_labels should be provided when num_class_embeds > 0"
+                )
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+
+        # pre-process
+        sample = self.conv_in(sample)
+        if pose_cond_fea is not None:
+            sample = sample + pose_cond_fea
+
+        # down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if (
+                hasattr(downsample_block, "has_cross_attention")
+                and downsample_block.has_cross_attention
+            ):
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=self.mode,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                )
+
+            down_block_res_samples += res_samples
+
+        if down_block_additional_residuals is not None:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_additional_residual = rearrange(
+                    down_block_additional_residual, "(b f) c h w -> b c f h w", f=down_block_res_sample.shape[2]
+                )
+                down_block_res_sample = (
+                    down_block_res_sample + down_block_additional_residual
+                )
+                new_down_block_res_samples += (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # mid
+        sample = self.mid_block(
+            sample,
+            emb,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            self_attention_additional_feats=self_attention_additional_feats,
+            mode=self.mode,
+        )
+
+        if mid_block_additional_residual is not None:
+            mid_block_additional_residual = rearrange(
+                mid_block_additional_residual, "(b f) c h w -> b c f h w", f=sample.shape[2]
+            )            
+            sample = sample + mid_block_additional_residual
+
+        # up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[
+                : -len(upsample_block.resnets)
+            ]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if (
+                hasattr(upsample_block, "has_cross_attention")
+                and upsample_block.has_cross_attention
+            ):
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=self.mode,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    encoder_hidden_states=encoder_hidden_states,
+                )
+
+        # post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet3DConditionOutput(sample=sample)
+
+    @classmethod
+    def from_pretrained_2d(
+        cls,
+        pretrained_model_path: PathLike,
+        motion_module_path: PathLike,
+        subfolder=None,
+        unet_additional_kwargs=None,
+        mm_zero_proj_out=False,
+    ):
+        pretrained_model_path = Path(pretrained_model_path)
+        motion_module_path = Path(motion_module_path)
+        if subfolder is not None:
+            pretrained_model_path = pretrained_model_path.joinpath(subfolder)
+        logger.info(
+            f"loaded temporal unet's pretrained weights from {pretrained_model_path} ..."
+        )
+
+        config_file = pretrained_model_path / "config.json"
+        if not (config_file.exists() and config_file.is_file()):
+            raise RuntimeError(f"{config_file} does not exist or is not a file")
+
+        unet_config = cls.load_config(config_file)
+        unet_config["_class_name"] = cls.__name__
+        unet_config["down_block_types"] = [
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ]
+        unet_config["up_block_types"] = [
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+        ]
+        unet_config["mid_block_type"] = "UNetMidBlock3DCrossAttn"
+
+        model = cls.from_config(unet_config, **unet_additional_kwargs)
+        # load the vanilla weights
+        # for key in unet_additional_kwargs['motion_module_kwargs'].keys():
+        #     print(key)
+        #     print(unet_additional_kwargs['motion_module_kwargs'][key])
+        #     print('__________________')
+        if pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME).exists():
+            logger.debug(
+                f"loading safeTensors weights from {pretrained_model_path} ..."
+            )
+            state_dict = load_file(
+                pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME), device="cpu"
+            )
+
+        elif pretrained_model_path.joinpath(WEIGHTS_NAME).exists():
+            logger.debug(f"loading weights from {pretrained_model_path} ...")
+            state_dict = torch.load(
+                pretrained_model_path.joinpath(WEIGHTS_NAME),
+                map_location="cpu",
+                weights_only=True,
+            )
+        else:
+            raise FileNotFoundError(f"no weights file found in {pretrained_model_path}")
+
+        # load the motion module weights
+        if motion_module_path.exists() and motion_module_path.is_file():
+            if motion_module_path.suffix.lower() in [".pth", ".pt", ".ckpt"]:
+                logger.info(f"Load motion module params from {motion_module_path}")
+                motion_state_dict = torch.load(
+                    motion_module_path, map_location="cpu", weights_only=True
+                )
+            elif motion_module_path.suffix.lower() == ".safetensors":
+                motion_state_dict = load_file(motion_module_path, device="cpu")
+            else:
+                raise RuntimeError(
+                    f"unknown file format for motion module weights: {motion_module_path.suffix}"
+                )
+            if mm_zero_proj_out:
+                logger.info(f"Zero initialize proj_out layers in motion module...")
+                new_motion_state_dict = OrderedDict()
+                for k in motion_state_dict:
+                    if "proj_out" in k:
+                        continue
+                    new_motion_state_dict[k] = motion_state_dict[k]
+                motion_state_dict = new_motion_state_dict
+
+            # merge the state dicts
+            state_dict.update(motion_state_dict)
+
+        # load the weights into the model
+        m, u = model.load_state_dict(state_dict, strict=False)
+        logger.debug(f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
+
+        params = [
+            p.numel() if "temporal" in n else 0 for n, p in model.named_parameters()
+        ]
+        logger.info(f"Loaded {sum(params) / 1e6}M-parameter motion module")
+
+        return model

src/models/unet_3d_blocks.py

@@ -0,0 +1,906 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+
+import pdb
+
+import torch
+from torch import nn
+
+from .motion_module import get_motion_module
+
+# from .motion_module import get_motion_module
+from .resnet import Downsample3D, ResnetBlock3D, Upsample3D
+from .transformer_3d import Transformer3DModel
+
+
+def get_down_block(
+    down_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    temb_channels,
+    add_downsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    downsample_padding=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    unet_use_cross_frame_attention=None,
+    unet_use_temporal_attention=None,
+    use_inflated_groupnorm=None,
+    use_motion_module=None,
+    motion_module_type=None,
+    motion_module_kwargs=None,
+    name_index=None,
+):
+    down_block_type = (
+        down_block_type[7:]
+        if down_block_type.startswith("UNetRes")
+        else down_block_type
+    )
+    if down_block_type == "DownBlock3D":
+        return DownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    elif down_block_type == "CrossAttnDownBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnDownBlock3D"
+            )
+        if name_index is not None:
+            name_index = f"CrossAttnDownBlock_{name_index}_"
+        return CrossAttnDownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+            unet_use_temporal_attention=unet_use_temporal_attention,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+            name=name_index,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    prev_output_channel,
+    temb_channels,
+    add_upsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    unet_use_cross_frame_attention=None,
+    unet_use_temporal_attention=None,
+    use_inflated_groupnorm=None,
+    use_motion_module=None,
+    motion_module_type=None,
+    motion_module_kwargs=None,
+    name_index=None,
+):
+    up_block_type = (
+        up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    )
+    if up_block_type == "UpBlock3D":
+        return UpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    elif up_block_type == "CrossAttnUpBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnUpBlock3D"
+            )
+        if name_index is not None:
+            name_index = f"CrossAttnUpBlock_{name_index}_"
+        return CrossAttnUpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+            unet_use_temporal_attention=unet_use_temporal_attention,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+            name=name_index,
+        )
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class UNetMidBlock3DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        output_scale_factor=1.0,
+        cross_attention_dim=1280,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+        name=None
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+        self.name = name
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock3D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+            )
+        ]
+        attentions = []
+        motion_modules = []
+        for i in range(num_layers):
+            if dual_cross_attention:
+                raise NotImplementedError
+            if self.name is not None:
+                attn_name = f"{self.name}_{i}_TransformerModel"
+            else:
+                attn_name = None
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    in_channels // attn_num_head_channels,
+                    in_channels=in_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                    name=attn_name,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=in_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+    def forward(
+        self,
+        hidden_states,
+        temb=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        self_attention_additional_feats=None,
+        mode=None,
+    ):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet, motion_module in zip(
+            self.attentions, self.resnets[1:], self.motion_modules
+        ):
+            hidden_states = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                self_attention_additional_feats=self_attention_additional_feats,
+                mode=mode,
+            ).sample
+            hidden_states = (
+                motion_module(
+                    hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                )
+                if motion_module is not None
+                else hidden_states
+            )
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class CrossAttnDownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        downsample_padding=1,
+        add_downsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+        name=None,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+        motion_modules = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        self.name=name
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            if self.name is not None:
+                attn_name = f"{self.name}_{i}_TransformerModel"
+            else:
+                attn_name = None
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                    name=attn_name,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        temb=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        self_attention_additional_feats=None,
+        mode=None,
+    ):
+        output_states = ()
+
+        for i, (resnet, attn, motion_module) in enumerate(
+            zip(self.resnets, self.attentions, self.motion_modules)
+        ):
+            # self.gradient_checkpointing = False
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                    self_attention_additional_feats,
+                    mode,
+                )[0]
+
+                # add motion module
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=mode,
+                ).sample
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_downsample=True,
+        downsample_padding=1,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        motion_modules = []
+
+        # use_motion_module = False
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
+        output_states = ()
+
+        for resnet, motion_module in zip(self.resnets, self.motion_modules):
+            # print(f"DownBlock3D {self.gradient_checkpointing = }")
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_motion_module=None,
+        use_inflated_groupnorm=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+        name=None
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+        motion_modules = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        self.name = name
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            if self.name is not None:
+                attn_name = f"{self.name}_{i}_TransformerModel"
+            else:
+                attn_name = None
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                    name=attn_name,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        res_hidden_states_tuple,
+        temb=None,
+        encoder_hidden_states=None,
+        upsample_size=None,
+        attention_mask=None,
+        self_attention_additional_feats=None,
+        mode=None,
+    ):
+        for i, (resnet, attn, motion_module) in enumerate(
+            zip(self.resnets, self.attentions, self.motion_modules)
+        ):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                    self_attention_additional_feats,
+                    mode,
+                )[0]
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    self_attention_additional_feats=self_attention_additional_feats,
+                    mode=mode,
+                ).sample
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states
+
+
+class UpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        motion_modules = []
+
+        # use_motion_module = False
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        res_hidden_states_tuple,
+        temb=None,
+        upsample_size=None,
+        encoder_hidden_states=None,
+    ):
+        for resnet, motion_module in zip(self.resnets, self.motion_modules):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            # print(f"UpBlock3D {self.gradient_checkpointing = }")
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states

src/pipelines/context.py

@@ -0,0 +1,76 @@
+# TODO: Adapted from cli
+from typing import Callable, List, Optional
+
+import numpy as np
+
+
+def ordered_halving(val):
+    bin_str = f"{val:064b}"
+    bin_flip = bin_str[::-1]
+    as_int = int(bin_flip, 2)
+
+    return as_int / (1 << 64)
+
+
+def uniform(
+    step: int = ...,
+    num_steps: Optional[int] = None,
+    num_frames: int = ...,
+    context_size: Optional[int] = None,
+    context_stride: int = 3,
+    context_overlap: int = 4,
+    closed_loop: bool = True,
+):
+    if num_frames <= context_size:
+        yield list(range(num_frames))
+        return
+
+    context_stride = min(
+        context_stride, int(np.ceil(np.log2(num_frames / context_size))) + 1
+    )
+
+    for context_step in 1 << np.arange(context_stride):
+        pad = int(round(num_frames * ordered_halving(step)))
+        for j in range(
+            int(ordered_halving(step) * context_step) + pad,
+            num_frames + pad + (0 if closed_loop else -context_overlap),
+            (context_size * context_step - context_overlap),
+        ):
+            yield [
+                e % num_frames
+                for e in range(j, j + context_size * context_step, context_step)
+            ]
+
+
+def get_context_scheduler(name: str) -> Callable:
+    if name == "uniform":
+        return uniform
+    else:
+        raise ValueError(f"Unknown context_overlap policy {name}")
+
+
+def get_total_steps(
+    scheduler,
+    timesteps: List[int],
+    num_steps: Optional[int] = None,
+    num_frames: int = ...,
+    context_size: Optional[int] = None,
+    context_stride: int = 3,
+    context_overlap: int = 4,
+    closed_loop: bool = True,
+):
+    return sum(
+        len(
+            list(
+                scheduler(
+                    i,
+                    num_steps,
+                    num_frames,
+                    context_size,
+                    context_stride,
+                    context_overlap,
+                )
+            )
+        )
+        for i in range(len(timesteps))
+    )

src/pipelines/pipeline_lmks2vid_long.py

@@ -0,0 +1,622 @@
+import inspect
+import math
+from dataclasses import dataclass
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+from diffusers import DiffusionPipeline
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers import (
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    EulerAncestralDiscreteScheduler,
+    EulerDiscreteScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+)
+from diffusers.utils import BaseOutput, deprecate, is_accelerate_available, logging
+from diffusers.utils.torch_utils import randn_tensor
+from einops import rearrange
+from tqdm import tqdm
+from transformers import CLIPImageProcessor
+
+from src.models.mutual_self_attention import ReferenceAttentionControl
+from src.pipelines.context import get_context_scheduler
+from src.pipelines.utils import get_tensor_interpolation_method
+
+
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(
+        dim=list(range(1, noise_pred_text.ndim)), keepdim=True
+    )
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = (
+        guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    )
+    print(f"{(std_text / std_cfg) = }")
+    return noise_cfg
+
+
+@dataclass
+class Pose2VideoPipelineOutput(BaseOutput):
+    videos: Union[torch.Tensor, np.ndarray]
+
+
+class Pose2VideoPipeline(DiffusionPipeline):
+    _optional_components = []
+
+    def __init__(
+        self,
+        vae,
+        image_encoder,
+        reference_unet,
+        denoising_unet,
+        pose_guider1,
+        pose_guider2,
+        scheduler: Union[
+            DDIMScheduler,
+            PNDMScheduler,
+            LMSDiscreteScheduler,
+            EulerDiscreteScheduler,
+            EulerAncestralDiscreteScheduler,
+            DPMSolverMultistepScheduler,
+        ],
+        audio_guider=None,
+        image_proj_model=None,
+        tokenizer=None,
+        text_encoder=None,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            image_encoder=image_encoder,
+            reference_unet=reference_unet,
+            denoising_unet=denoising_unet,
+            pose_guider1=pose_guider1,
+            pose_guider2=pose_guider2,
+            scheduler=scheduler,
+            audio_guider=audio_guider,
+            image_proj_model=image_proj_model,
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.clip_image_processor = CLIPImageProcessor()
+        self.ref_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True
+        )
+        self.cond_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor,
+            do_convert_rgb=True,
+            do_normalize=False,
+        )
+
+    def enable_vae_slicing(self):
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    def _execution_device(self):
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def decode_latents(self, latents):
+        video_length = latents.shape[2]
+        latents = 1 / 0.18215 * latents
+        latents = rearrange(latents, "b c f h w -> (b f) c h w")
+        # video = self.vae.decode(latents).sample
+        video = []
+        for frame_idx in tqdm(range(latents.shape[0])):
+            video.append(self.vae.decode(latents[frame_idx : frame_idx + 1]).sample)
+        video = torch.cat(video)
+        video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
+        video = (video / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        video = video.cpu().float().numpy()
+        return video
+
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        width,
+        height,
+        video_length,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            video_length,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(
+                shape, generator=generator, device=device, dtype=dtype
+            )
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def _encode_prompt(
+        self,
+        prompt,
+        device,
+        num_videos_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt,
+    ):
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(
+            prompt, padding="longest", return_tensors="pt"
+        ).input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+            text_input_ids, untruncated_ids
+        ):
+            removed_text = self.tokenizer.batch_decode(
+                untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+            )
+
+        if (
+            hasattr(self.text_encoder.config, "use_attention_mask")
+            and self.text_encoder.config.use_attention_mask
+        ):
+            attention_mask = text_inputs.attention_mask.to(device)
+        else:
+            attention_mask = None
+
+        text_embeddings = self.text_encoder(
+            text_input_ids.to(device),
+            attention_mask=attention_mask,
+        )
+        text_embeddings = text_embeddings[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_videos_per_prompt, 1)
+        text_embeddings = text_embeddings.view(
+            bs_embed * num_videos_per_prompt, seq_len, -1
+        )
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if (
+                hasattr(self.text_encoder.config, "use_attention_mask")
+                and self.text_encoder.config.use_attention_mask
+            ):
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            uncond_embeddings = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            uncond_embeddings = uncond_embeddings[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_videos_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(
+                batch_size * num_videos_per_prompt, seq_len, -1
+            )
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        return text_embeddings
+
+    def interpolate_latents(
+        self, latents: torch.Tensor, interpolation_factor: int, device
+    ):
+        if interpolation_factor < 2:
+            return latents
+
+        new_latents = torch.zeros(
+            (
+                latents.shape[0],
+                latents.shape[1],
+                ((latents.shape[2] - 1) * interpolation_factor) + 1,
+                latents.shape[3],
+                latents.shape[4],
+            ),
+            device=latents.device,
+            dtype=latents.dtype,
+        )
+
+        org_video_length = latents.shape[2]
+        rate = [i / interpolation_factor for i in range(interpolation_factor)][1:]
+
+        new_index = 0
+
+        v0 = None
+        v1 = None
+
+        for i0, i1 in zip(range(org_video_length), range(org_video_length)[1:]):
+            v0 = latents[:, :, i0, :, :]
+            v1 = latents[:, :, i1, :, :]
+
+            new_latents[:, :, new_index, :, :] = v0
+            new_index += 1
+
+            for f in rate:
+                v = get_tensor_interpolation_method()(
+                    v0.to(device=device), v1.to(device=device), f
+                )
+                new_latents[:, :, new_index, :, :] = v.to(latents.device)
+                new_index += 1
+
+        new_latents[:, :, new_index, :, :] = v1
+        new_index += 1
+
+        return new_latents
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        ref_image,
+        pose_up_images,
+        pose_down_images, 
+        width,
+        height,
+        video_length,
+        num_inference_steps,
+        guidance_scale,
+        audio_features=None, 
+        num_images_per_prompt=1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: Optional[str] = "tensor",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        guidance_rescale=0.,  # 0.7
+        **kwargs,
+    ):
+        # Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        device = self._execution_device
+
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        batch_size = 1
+
+        # Prepare clip image embeds
+        # NOTE: 这里是否需要 resize 到 (224, 224) 需要观察
+        clip_image = self.clip_image_processor.preprocess(
+            ref_image.resize((224, 224)), return_tensors="pt"
+        ).pixel_values
+        # If image_proj_model is not None, means enable ip-adapter
+        if self.image_proj_model is not None:
+            clip_image_embeds = self.image_encoder(
+                clip_image.to(device, dtype=self.image_encoder.dtype),
+                output_hidden_states=True,
+            ).hidden_states[-2]
+            image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+            uncond_image_prompt_embeds = self.image_proj_model(
+                torch.zeros_like(clip_image_embeds)
+            )
+            text_prompt_embeds = self._encode_prompt(
+                "best quality, high quality",
+                device,
+                num_images_per_prompt,
+                do_classifier_free_guidance,
+                negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
+            )
+            # Concat image and text embeddings
+            if do_classifier_free_guidance:
+                (
+                    uncond_text_prompt_embeds,
+                    text_prompt_embeds,
+                ) = text_prompt_embeds.chunk(2)
+                uncond_encoder_hidden_states = torch.cat(
+                    [uncond_text_prompt_embeds, uncond_image_prompt_embeds], dim=1
+                )
+            encoder_hidden_states = torch.cat(
+                [text_prompt_embeds, image_prompt_embeds], dim=1
+            )
+
+        else:
+            clip_image_embeds = self.image_encoder(
+                clip_image.to(device, dtype=self.image_encoder.dtype)
+            ).image_embeds
+            encoder_hidden_states = clip_image_embeds.unsqueeze(1)
+            uncond_encoder_hidden_states = torch.zeros_like(encoder_hidden_states)
+
+        if do_classifier_free_guidance:
+            encoder_hidden_states = torch.cat(
+                [uncond_encoder_hidden_states, encoder_hidden_states], dim=0
+            )
+
+        num_channels_latents = self.denoising_unet.in_channels
+        
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            width,
+            height,
+            video_length,
+            clip_image_embeds.dtype,
+            device,
+            generator,
+        )
+
+        # Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # Prepare ref image latents
+        ref_image_tensor = self.ref_image_processor.preprocess(
+            ref_image, height=height, width=width
+        )  # (bs, c, width, height)
+        ref_image_tensor = ref_image_tensor.to(
+            dtype=self.vae.dtype, device=self.vae.device
+        )
+        ref_image_latents = self.vae.encode(ref_image_tensor).latent_dist.mean
+        ref_image_latents = ref_image_latents * 0.18215  # (b, 4, h, w)
+        ref_image_latents = ref_image_latents.unsqueeze(2)
+
+        # Prepare a list of pose condition images
+        pose_up_cond_tensor_list, pose_down_cond_tensor_list = [], []
+        for i, pose_up_image in enumerate(pose_up_images):
+            pose_up_cond_tensor = self.cond_image_processor.preprocess(
+                pose_up_image, height=height, width=width
+            )
+            pose_up_cond_tensor = pose_up_cond_tensor.unsqueeze(2)  # (bs, c, 1, h, w)
+            pose_up_cond_tensor_list.append(pose_up_cond_tensor)
+            pose_down_cond_tensor = self.cond_image_processor.preprocess(
+                pose_down_images[i], height=height, width=width
+            )
+            pose_down_cond_tensor = pose_down_cond_tensor.unsqueeze(2)  # (bs, c, 1, h, w)
+            pose_down_cond_tensor_list.append(pose_down_cond_tensor)
+        pose_up_cond_tensor = torch.cat(pose_up_cond_tensor_list, dim=2)  # (bs, c, t, h, w)
+        pose_up_cond_tensor = pose_up_cond_tensor.to(
+            device=device, dtype=self.pose_guider1.dtype
+        )
+        pose_down_cond_tensor = torch.cat(pose_down_cond_tensor_list, dim=2)  # (bs, c, t, h, w)
+        pose_up_fea = self.pose_guider1(pose_up_cond_tensor)
+        pose_down_cond_tensor = pose_down_cond_tensor.to(device=device, dtype=self.pose_guider2.dtype)
+        pose_down_fea = self.pose_guider2(pose_down_cond_tensor)
+        pose_fea = pose_up_fea + pose_down_fea
+        
+        context_schedule = "uniform"
+        context_frames = 24
+        context_stride = 1
+        context_overlap = 4  # 4
+        context_batch_size = 1
+        context_scheduler = get_context_scheduler(context_schedule)
+
+        # denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        middle_results = []
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            self_attention_additional_feats = {}
+            for i, t in enumerate(timesteps):
+                noise_pred = torch.zeros(
+                    (
+                        latents.shape[0] * (2 if do_classifier_free_guidance else 1),
+                        *latents.shape[1:],
+                    ),
+                    device=latents.device,
+                    dtype=latents.dtype,
+                )
+                counter = torch.zeros(
+                    (1, 1, latents.shape[2], 1, 1),
+                    device=latents.device,
+                    dtype=latents.dtype,
+                )
+
+                # 1. Forward reference image
+                if i == 0:
+                    self.reference_unet(
+                        # ref_image_latents.repeat(
+                            # (2 if do_classifier_free_guidance else 1), 1, 1, 1
+                        # ),
+                        torch.cat([torch.zeros_like(ref_image_latents), ref_image_latents]) \
+                        if do_classifier_free_guidance else \
+                        ref_image_latents,
+                        torch.zeros_like(t),
+                        # t,
+                        encoder_hidden_states=encoder_hidden_states,
+                        self_attention_additional_feats=self_attention_additional_feats,
+                        return_dict=False,
+                    )
+                context_queue = list(
+                    context_scheduler(
+                        0,
+                        num_inference_steps,
+                        latents.shape[2],
+                        context_frames,
+                        context_stride,
+                        context_overlap,
+                    )
+                )
+
+                num_context_batches = math.ceil(len(context_queue) / context_batch_size)
+                global_context = []
+                for i in range(num_context_batches):
+                    global_context.append(
+                        context_queue[
+                            i * context_batch_size : (i + 1) * context_batch_size
+                        ]
+                    )
+
+                for context in global_context:
+                    # 3.1 expand the latents if we are doing classifier free guidance
+                    latent_model_input = (
+                        torch.cat([latents[:, :, c] for c in context])
+                        .to(device)
+                        .repeat(2 if do_classifier_free_guidance else 1, 1, 1, 1, 1)
+                    )
+                    latent_model_input = self.scheduler.scale_model_input(
+                        latent_model_input, t
+                    )
+                    b, c, f, h, w = latent_model_input.shape
+                    latent_pose_input = torch.cat(
+                        [pose_fea[:, :, c] for c in context]
+                    ).repeat(2 if do_classifier_free_guidance else 1, 1, 1, 1, 1)
+
+                    pred = self.denoising_unet(
+                        latent_model_input,
+                        t,
+                        encoder_hidden_states=encoder_hidden_states[:b],
+                        pose_cond_fea=latent_pose_input,
+                        self_attention_additional_feats=self_attention_additional_feats,
+                        return_dict=False,
+                    )[0]
+
+                    for j, c in enumerate(context):
+                        noise_pred[:, :, c] = noise_pred[:, :, c] + pred
+                        counter[:, :, c] = counter[:, :, c] + 1
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = (noise_pred / counter).chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (
+                        noise_pred_text - noise_pred_uncond
+                    )
+
+                if do_classifier_free_guidance and guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(
+                        noise_pred,
+                        noise_pred_text,
+                        guidance_rescale=guidance_rescale,
+                    )
+
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs
+                ).prev_sample
+
+                if i == len(timesteps) - 1 or (
+                    (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
+                ):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+
+        interpolation_factor = 1
+        latents = self.interpolate_latents(latents, interpolation_factor, device)
+        # Post-processing
+        images = self.decode_latents(latents)  # (b, c, f, h, w)
+
+        # Convert to tensor
+        if output_type == "tensor":
+            images = torch.from_numpy(images)
+
+        if not return_dict:
+            return images
+
+        return Pose2VideoPipelineOutput(videos=images)

src/pipelines/pipeline_pose2img.py

@@ -0,0 +1,360 @@
+import inspect
+from dataclasses import dataclass
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+from diffusers import DiffusionPipeline
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers import (
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    EulerAncestralDiscreteScheduler,
+    EulerDiscreteScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+)
+from diffusers.utils import BaseOutput, is_accelerate_available
+from diffusers.utils.torch_utils import randn_tensor
+from einops import rearrange
+from tqdm import tqdm
+from transformers import CLIPImageProcessor
+
+from src.models.mutual_self_attention import ReferenceAttentionControl
+
+
+@dataclass
+class Pose2ImagePipelineOutput(BaseOutput):
+    images: Union[torch.Tensor, np.ndarray]
+
+
+class Pose2ImagePipeline(DiffusionPipeline):
+    _optional_components = []
+
+    def __init__(
+        self,
+        vae,
+        image_encoder,
+        reference_unet,
+        denoising_unet,
+        pose_guider,
+        scheduler: Union[
+            DDIMScheduler,
+            PNDMScheduler,
+            LMSDiscreteScheduler,
+            EulerDiscreteScheduler,
+            EulerAncestralDiscreteScheduler,
+            DPMSolverMultistepScheduler,
+        ],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            image_encoder=image_encoder,
+            reference_unet=reference_unet,
+            denoising_unet=denoising_unet,
+            pose_guider=pose_guider,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.clip_image_processor = CLIPImageProcessor()
+        self.ref_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True
+        )
+        self.cond_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor,
+            do_convert_rgb=True,
+            do_normalize=False,
+        )
+
+    def enable_vae_slicing(self):
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    def _execution_device(self):
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def decode_latents(self, latents):
+        video_length = latents.shape[2]
+        latents = 1 / 0.18215 * latents
+        latents = rearrange(latents, "b c f h w -> (b f) c h w")
+        # video = self.vae.decode(latents).sample
+        video = []
+        for frame_idx in tqdm(range(latents.shape[0])):
+            video.append(self.vae.decode(latents[frame_idx : frame_idx + 1]).sample)
+        video = torch.cat(video)
+        video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
+        video = (video / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        video = video.cpu().float().numpy()
+        return video
+
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        width,
+        height,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(
+                shape, generator=generator, device=device, dtype=dtype
+            )
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def prepare_condition(
+        self,
+        cond_image,
+        width,
+        height,
+        device,
+        dtype,
+        do_classififer_free_guidance=False,
+    ):
+        image = self.cond_image_processor.preprocess(
+            cond_image, height=height, width=width
+        ).to(dtype=torch.float32)
+
+        image = image.to(device=device, dtype=dtype)
+
+        if do_classififer_free_guidance:
+            image = torch.cat([image] * 2)
+
+        return image
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        ref_image,
+        pose_image,
+        width,
+        height,
+        num_inference_steps,
+        guidance_scale,
+        num_images_per_prompt=1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: Optional[str] = "tensor",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        # Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        device = self._execution_device
+
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        batch_size = 1
+
+        # Prepare clip image embeds
+        clip_image = self.clip_image_processor.preprocess(
+            ref_image.resize((224, 224)), return_tensors="pt"
+        ).pixel_values
+        clip_image_embeds = self.image_encoder(
+            clip_image.to(device, dtype=self.image_encoder.dtype)
+        ).image_embeds
+        image_prompt_embeds = clip_image_embeds.unsqueeze(1)
+        uncond_image_prompt_embeds = torch.zeros_like(image_prompt_embeds)
+
+        if do_classifier_free_guidance:
+            image_prompt_embeds = torch.cat(
+                [uncond_image_prompt_embeds, image_prompt_embeds], dim=0
+            )
+
+        reference_control_writer = ReferenceAttentionControl(
+            self.reference_unet,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            mode="write",
+            batch_size=batch_size,
+            fusion_blocks="full",
+        )
+        reference_control_reader = ReferenceAttentionControl(
+            self.denoising_unet,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            mode="read",
+            batch_size=batch_size,
+            fusion_blocks="full",
+        )
+
+        num_channels_latents = self.denoising_unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            width,
+            height,
+            clip_image_embeds.dtype,
+            device,
+            generator,
+        )
+        latents = latents.unsqueeze(2)  # (bs, c, 1, h', w')
+        latents_dtype = latents.dtype
+
+        # Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # Prepare ref image latents
+        ref_image_tensor = self.ref_image_processor.preprocess(
+            ref_image, height=height, width=width
+        )  # (bs, c, width, height)
+        ref_image_tensor = ref_image_tensor.to(
+            dtype=self.vae.dtype, device=self.vae.device
+        )
+        ref_image_latents = self.vae.encode(ref_image_tensor).latent_dist.mean
+        ref_image_latents = ref_image_latents * 0.18215  # (b, 4, h, w)
+
+        # Prepare pose condition image
+        pose_cond_tensor = self.cond_image_processor.preprocess(
+            pose_image, height=height, width=width
+        )
+        pose_cond_tensor = pose_cond_tensor.unsqueeze(2)  # (bs, c, 1, h, w)
+        pose_cond_tensor = pose_cond_tensor.to(
+            device=device, dtype=self.pose_guider.dtype
+        )
+        pose_fea = self.pose_guider(pose_cond_tensor)
+        pose_fea = (
+            torch.cat([pose_fea] * 2) if do_classifier_free_guidance else pose_fea
+        )
+
+        # denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # 1. Forward reference image
+                if i == 0:
+                    self.reference_unet(
+                        ref_image_latents.repeat(
+                            (2 if do_classifier_free_guidance else 1), 1, 1, 1
+                        ),
+                        torch.zeros_like(t),
+                        encoder_hidden_states=image_prompt_embeds,
+                        return_dict=False,
+                    )
+
+                    # 2. Update reference unet feature into denosing net
+                    reference_control_reader.update(reference_control_writer)
+
+                # 3.1 expand the latents if we are doing classifier free guidance
+                latent_model_input = (
+                    torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                )
+                latent_model_input = self.scheduler.scale_model_input(
+                    latent_model_input, t
+                )
+
+                noise_pred = self.denoising_unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=image_prompt_embeds,
+                    pose_cond_fea=pose_fea,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (
+                        noise_pred_text - noise_pred_uncond
+                    )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs, return_dict=False
+                )[0]
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or (
+                    (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
+                ):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+            reference_control_reader.clear()
+            reference_control_writer.clear()
+
+        # Post-processing
+        image = self.decode_latents(latents)  # (b, c, 1, h, w)
+
+        # Convert to tensor
+        if output_type == "tensor":
+            image = torch.from_numpy(image)
+
+        if not return_dict:
+            return image
+
+        return Pose2ImagePipelineOutput(images=image)

src/pipelines/pipeline_pose2vid.py

@@ -0,0 +1,458 @@
+import inspect
+from dataclasses import dataclass
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+from diffusers import DiffusionPipeline
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers import (DDIMScheduler, DPMSolverMultistepScheduler,
+                                  EulerAncestralDiscreteScheduler,
+                                  EulerDiscreteScheduler, LMSDiscreteScheduler,
+                                  PNDMScheduler)
+from diffusers.utils import BaseOutput, is_accelerate_available
+from diffusers.utils.torch_utils import randn_tensor
+from einops import rearrange
+from tqdm import tqdm
+from transformers import CLIPImageProcessor
+
+from src.models.mutual_self_attention import ReferenceAttentionControl
+
+
+@dataclass
+class Pose2VideoPipelineOutput(BaseOutput):
+    videos: Union[torch.Tensor, np.ndarray]
+
+
+class Pose2VideoPipeline(DiffusionPipeline):
+    _optional_components = []
+
+    def __init__(
+        self,
+        vae,
+        image_encoder,
+        reference_unet,
+        denoising_unet,
+        pose_guider,
+        scheduler: Union[
+            DDIMScheduler,
+            PNDMScheduler,
+            LMSDiscreteScheduler,
+            EulerDiscreteScheduler,
+            EulerAncestralDiscreteScheduler,
+            DPMSolverMultistepScheduler,
+        ],
+        image_proj_model=None,
+        tokenizer=None,
+        text_encoder=None,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            image_encoder=image_encoder,
+            reference_unet=reference_unet,
+            denoising_unet=denoising_unet,
+            pose_guider=pose_guider,
+            scheduler=scheduler,
+            image_proj_model=image_proj_model,
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.clip_image_processor = CLIPImageProcessor()
+        self.ref_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True
+        )
+        self.cond_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor,
+            do_convert_rgb=True,
+            do_normalize=False,
+        )
+
+    def enable_vae_slicing(self):
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    def _execution_device(self):
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def decode_latents(self, latents):
+        video_length = latents.shape[2]
+        latents = 1 / 0.18215 * latents
+        latents = rearrange(latents, "b c f h w -> (b f) c h w")
+        # video = self.vae.decode(latents).sample
+        video = []
+        for frame_idx in tqdm(range(latents.shape[0])):
+            video.append(self.vae.decode(latents[frame_idx : frame_idx + 1]).sample)
+        video = torch.cat(video)
+        video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
+        video = (video / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        video = video.cpu().float().numpy()
+        return video
+
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        width,
+        height,
+        video_length,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            video_length,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(
+                shape, generator=generator, device=device, dtype=dtype
+            )
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def _encode_prompt(
+        self,
+        prompt,
+        device,
+        num_videos_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt,
+    ):
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(
+            prompt, padding="longest", return_tensors="pt"
+        ).input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+            text_input_ids, untruncated_ids
+        ):
+            removed_text = self.tokenizer.batch_decode(
+                untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+            )
+
+        if (
+            hasattr(self.text_encoder.config, "use_attention_mask")
+            and self.text_encoder.config.use_attention_mask
+        ):
+            attention_mask = text_inputs.attention_mask.to(device)
+        else:
+            attention_mask = None
+
+        text_embeddings = self.text_encoder(
+            text_input_ids.to(device),
+            attention_mask=attention_mask,
+        )
+        text_embeddings = text_embeddings[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_videos_per_prompt, 1)
+        text_embeddings = text_embeddings.view(
+            bs_embed * num_videos_per_prompt, seq_len, -1
+        )
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if (
+                hasattr(self.text_encoder.config, "use_attention_mask")
+                and self.text_encoder.config.use_attention_mask
+            ):
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            uncond_embeddings = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            uncond_embeddings = uncond_embeddings[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_videos_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(
+                batch_size * num_videos_per_prompt, seq_len, -1
+            )
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        return text_embeddings
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        ref_image,
+        pose_images,
+        width,
+        height,
+        video_length,
+        num_inference_steps,
+        guidance_scale,
+        num_images_per_prompt=1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: Optional[str] = "tensor",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        # Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        device = self._execution_device
+
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        batch_size = 1
+
+        # Prepare clip image embeds
+        clip_image = self.clip_image_processor.preprocess(
+            ref_image, return_tensors="pt"
+        ).pixel_values
+        clip_image_embeds = self.image_encoder(
+            clip_image.to(device, dtype=self.image_encoder.dtype)
+        ).image_embeds
+        encoder_hidden_states = clip_image_embeds.unsqueeze(1)
+        uncond_encoder_hidden_states = torch.zeros_like(encoder_hidden_states)
+
+        if do_classifier_free_guidance:
+            encoder_hidden_states = torch.cat(
+                [uncond_encoder_hidden_states, encoder_hidden_states], dim=0
+            )
+        reference_control_writer = ReferenceAttentionControl(
+            self.reference_unet,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            mode="write",
+            batch_size=batch_size,
+            fusion_blocks="full",
+        )
+        reference_control_reader = ReferenceAttentionControl(
+            self.denoising_unet,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            mode="read",
+            batch_size=batch_size,
+            fusion_blocks="full",
+        )
+
+        num_channels_latents = self.denoising_unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            width,
+            height,
+            video_length,
+            clip_image_embeds.dtype,
+            device,
+            generator,
+        )
+
+        # Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # Prepare ref image latents
+        ref_image_tensor = self.ref_image_processor.preprocess(
+            ref_image, height=height, width=width
+        )  # (bs, c, width, height)
+        ref_image_tensor = ref_image_tensor.to(
+            dtype=self.vae.dtype, device=self.vae.device
+        )
+        ref_image_latents = self.vae.encode(ref_image_tensor).latent_dist.mean
+        ref_image_latents = ref_image_latents * 0.18215  # (b, 4, h, w)
+
+        # Prepare a list of pose condition images
+        pose_cond_tensor_list = []
+        for pose_image in pose_images:
+            pose_cond_tensor = (
+                torch.from_numpy(np.array(pose_image.resize((width, height)))) / 255.0
+            )
+            pose_cond_tensor = pose_cond_tensor.permute(2, 0, 1).unsqueeze(
+                1
+            )  # (c, 1, h, w)
+            pose_cond_tensor_list.append(pose_cond_tensor)
+        pose_cond_tensor = torch.cat(pose_cond_tensor_list, dim=1)  # (c, t, h, w)
+        pose_cond_tensor = pose_cond_tensor.unsqueeze(0)
+        pose_cond_tensor = pose_cond_tensor.to(
+            device=device, dtype=self.pose_guider.dtype
+        )
+        pose_fea = self.pose_guider(pose_cond_tensor)
+        pose_fea = (
+            torch.cat([pose_fea] * 2) if do_classifier_free_guidance else pose_fea
+        )
+
+        # denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # 1. Forward reference image
+                if i == 0:
+                    self.reference_unet(
+                        ref_image_latents.repeat(
+                            (2 if do_classifier_free_guidance else 1), 1, 1, 1
+                        ),
+                        torch.zeros_like(t),
+                        # t,
+                        encoder_hidden_states=encoder_hidden_states,
+                        return_dict=False,
+                    )
+                    reference_control_reader.update(reference_control_writer)
+
+                # 3.1 expand the latents if we are doing classifier free guidance
+                latent_model_input = (
+                    torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                )
+                latent_model_input = self.scheduler.scale_model_input(
+                    latent_model_input, t
+                )
+
+                noise_pred = self.denoising_unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=encoder_hidden_states,
+                    pose_cond_fea=pose_fea,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (
+                        noise_pred_text - noise_pred_uncond
+                    )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs, return_dict=False
+                )[0]
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or (
+                    (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
+                ):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+            reference_control_reader.clear()
+            reference_control_writer.clear()
+
+        # Post-processing
+        images = self.decode_latents(latents)  # (b, c, f, h, w)
+
+        # Convert to tensor
+        if output_type == "tensor":
+            images = torch.from_numpy(images)
+
+        if not return_dict:
+            return images
+
+        return Pose2VideoPipelineOutput(videos=images)