# Modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py#L148 # Copyright 2025 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import inspect from PIL import Image import PIL from typing import Any, Callable, Dict, List, Optional, Union from .processor import JointAttnProcessor2_0 from diffusers import StableDiffusion3Img2ImgPipeline import torch from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback from diffusers.image_processor import PipelineImageInput, VaeImageProcessor from diffusers.loaders import FromSingleFileMixin, SD3IPAdapterMixin, SD3LoraLoaderMixin from diffusers.models.autoencoders import AutoencoderKL from diffusers.models.transformers import SD3Transformer2DModel from diffusers.schedulers import FlowMatchEulerDiscreteScheduler from diffusers.utils import ( USE_PEFT_BACKEND, is_torch_xla_available, logging, replace_example_docstring, scale_lora_layers, unscale_lora_layers, ) logger = logging.get_logger(__name__) from diffusers.utils.torch_utils import randn_tensor from diffusers.pipelines.pipeline_utils import DiffusionPipeline from diffusers.pipelines.stable_diffusion_3.pipeline_output import StableDiffusion3PipelineOutput from diffusers.pipelines.stable_diffusion_3 import StableDiffusion3Pipeline from .sd3_pipeline import VSFStableDiffusion3Pipeline # Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift def calculate_shift( image_seq_len, base_seq_len: int = 256, max_seq_len: int = 4096, base_shift: float = 0.5, max_shift: float = 1.15, ): m = (max_shift - base_shift) / (max_seq_len - base_seq_len) b = base_shift - m * base_seq_len mu = image_seq_len * m + b return mu if is_torch_xla_available(): import torch_xla.core.xla_model as xm XLA_AVAILABLE = True else: XLA_AVAILABLE = False def retrieve_latents( encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" ): if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": return encoder_output.latent_dist.sample(generator) elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": return encoder_output.latent_dist.mode() elif hasattr(encoder_output, "latents"): return encoder_output.latents else: raise AttributeError("Could not access latents of provided encoder_output") # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps def retrieve_timesteps( scheduler, num_inference_steps: Optional[int] = None, device: Optional[Union[str, torch.device]] = None, timesteps: Optional[List[int]] = None, sigmas: Optional[List[float]] = None, **kwargs, ): r""" Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. Args: scheduler (`SchedulerMixin`): The scheduler to get timesteps from. num_inference_steps (`int`): The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` must be `None`. device (`str` or `torch.device`, *optional*): The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. timesteps (`List[int]`, *optional*): Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, `num_inference_steps` and `sigmas` must be `None`. sigmas (`List[float]`, *optional*): Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, `num_inference_steps` and `timesteps` must be `None`. Returns: `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the second element is the number of inference steps. """ if timesteps is not None and sigmas is not None: raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") if timesteps is not None: accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) if not accepts_timesteps: raise ValueError( f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" f" timestep schedules. Please check whether you are using the correct scheduler." ) scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) timesteps = scheduler.timesteps num_inference_steps = len(timesteps) elif sigmas is not None: accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) if not accept_sigmas: raise ValueError( f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" f" sigmas schedules. Please check whether you are using the correct scheduler." ) scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) timesteps = scheduler.timesteps num_inference_steps = len(timesteps) else: scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) timesteps = scheduler.timesteps return timesteps, num_inference_steps class VSFStableDiffusion3Img2ImgPipeline(VSFStableDiffusion3Pipeline): def check_inputs( self, prompt, prompt_2, prompt_3, height, width, strength, negative_prompt=None, negative_prompt_2=None, negative_prompt_3=None, prompt_embeds=None, negative_prompt_embeds=None, pooled_prompt_embeds=None, negative_pooled_prompt_embeds=None, callback_on_step_end_tensor_inputs=None, max_sequence_length=None, ): if ( height % (self.vae_scale_factor * self.patch_size) != 0 or width % (self.vae_scale_factor * self.patch_size) != 0 ): raise ValueError( f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}." f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}." ) if strength < 0 or strength > 1: raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") if callback_on_step_end_tensor_inputs is not None and not all( k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs ): raise ValueError( f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" ) if prompt is not None and prompt_embeds is not None: raise ValueError( f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" " only forward one of the two." ) elif prompt_2 is not None and prompt_embeds is not None: raise ValueError( f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" " only forward one of the two." ) elif prompt_3 is not None and prompt_embeds is not None: raise ValueError( f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" " only forward one of the two." ) elif prompt is None and prompt_embeds is None: raise ValueError( "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." ) elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") if negative_prompt is not None and negative_prompt_embeds is not None: raise ValueError( f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" f" {negative_prompt_embeds}. Please make sure to only forward one of the two." ) elif negative_prompt_2 is not None and negative_prompt_embeds is not None: raise ValueError( f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" f" {negative_prompt_embeds}. Please make sure to only forward one of the two." ) elif negative_prompt_3 is not None and negative_prompt_embeds is not None: raise ValueError( f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:" f" {negative_prompt_embeds}. Please make sure to only forward one of the two." ) if prompt_embeds is not None and negative_prompt_embeds is not None: if prompt_embeds.shape != negative_prompt_embeds.shape: raise ValueError( "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" f" {negative_prompt_embeds.shape}." ) if prompt_embeds is not None and pooled_prompt_embeds is None: raise ValueError( "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." ) if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: raise ValueError( "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." ) if max_sequence_length is not None and max_sequence_length > 512: raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") def get_timesteps(self, num_inference_steps, strength, device): # get the original timestep using init_timestep init_timestep = min(num_inference_steps * strength, num_inference_steps) t_start = int(max(num_inference_steps - init_timestep, 0)) timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] if hasattr(self.scheduler, "set_begin_index"): self.scheduler.set_begin_index(t_start * self.scheduler.order) return timesteps, num_inference_steps - t_start def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): raise ValueError( f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" ) image = image.to(device=device, dtype=dtype) batch_size = batch_size * num_images_per_prompt if image.shape[1] == self.vae.config.latent_channels: init_latents = image else: 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." ) elif isinstance(generator, list): init_latents = [ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) for i in range(batch_size) ] init_latents = torch.cat(init_latents, dim=0) else: init_latents = retrieve_latents(self.vae.encode(image), generator=generator) init_latents = (init_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: # expand init_latents for batch_size additional_image_per_prompt = batch_size // init_latents.shape[0] init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: raise ValueError( f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." ) else: init_latents = torch.cat([init_latents], dim=0) shape = init_latents.shape noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) # get latents init_latents = self.scheduler.scale_noise(init_latents, timestep, noise) latents = init_latents.to(device=device, dtype=dtype) return latents @property def guidance_scale(self): return self._guidance_scale @property def joint_attention_kwargs(self): return self._joint_attention_kwargs @property def clip_skip(self): return self._clip_skip # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` # corresponds to doing no classifier free guidance. @property def do_classifier_free_guidance(self): return self._guidance_scale > 1 @property def num_timesteps(self): return self._num_timesteps @property def interrupt(self): return self._interrupt # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_image def encode_image(self, image: PipelineImageInput, device: torch.device) -> torch.Tensor: """Encodes the given image into a feature representation using a pre-trained image encoder. Args: image (`PipelineImageInput`): Input image to be encoded. device: (`torch.device`): Torch device. Returns: `torch.Tensor`: The encoded image feature representation. """ if not isinstance(image, torch.Tensor): image = self.feature_extractor(image, return_tensors="pt").pixel_values image = image.to(device=device, dtype=self.dtype) return self.image_encoder(image, output_hidden_states=True).hidden_states[-2] # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.prepare_ip_adapter_image_embeds def prepare_ip_adapter_image_embeds( self, ip_adapter_image: Optional[PipelineImageInput] = None, ip_adapter_image_embeds: Optional[torch.Tensor] = None, device: Optional[torch.device] = None, num_images_per_prompt: int = 1, do_classifier_free_guidance: bool = True, ) -> torch.Tensor: """Prepares image embeddings for use in the IP-Adapter. Either `ip_adapter_image` or `ip_adapter_image_embeds` must be passed. Args: ip_adapter_image (`PipelineImageInput`, *optional*): The input image to extract features from for IP-Adapter. ip_adapter_image_embeds (`torch.Tensor`, *optional*): Precomputed image embeddings. device: (`torch.device`, *optional*): Torch device. num_images_per_prompt (`int`, defaults to 1): Number of images that should be generated per prompt. do_classifier_free_guidance (`bool`, defaults to True): Whether to use classifier free guidance or not. """ device = device or self._execution_device if ip_adapter_image_embeds is not None: if do_classifier_free_guidance: single_negative_image_embeds, single_image_embeds = ip_adapter_image_embeds.chunk(2) else: single_image_embeds = ip_adapter_image_embeds elif ip_adapter_image is not None: single_image_embeds = self.encode_image(ip_adapter_image, device) if do_classifier_free_guidance: single_negative_image_embeds = torch.zeros_like(single_image_embeds) else: raise ValueError("Neither `ip_adapter_image_embeds` or `ip_adapter_image_embeds` were provided.") image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) if do_classifier_free_guidance: negative_image_embeds = torch.cat([single_negative_image_embeds] * num_images_per_prompt, dim=0) image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) return image_embeds.to(device=device) # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.enable_sequential_cpu_offload def enable_sequential_cpu_offload(self, *args, **kwargs): if self.image_encoder is not None and "image_encoder" not in self._exclude_from_cpu_offload: logger.warning( "`pipe.enable_sequential_cpu_offload()` might fail for `image_encoder` if it uses " "`torch.nn.MultiheadAttention`. You can exclude `image_encoder` from CPU offloading by calling " "`pipe._exclude_from_cpu_offload.append('image_encoder')` before `pipe.enable_sequential_cpu_offload()`." ) super().enable_sequential_cpu_offload(*args, **kwargs) @torch.no_grad() def __call__( self, prompt: Union[str, List[str]] = None, prompt_2: Optional[Union[str, List[str]]] = None, prompt_3: Optional[Union[str, List[str]]] = None, height: Optional[int] = None, width: Optional[int] = None, image: PipelineImageInput = None, strength: float = 0.6, num_inference_steps: int = 50, sigmas: Optional[List[float]] = None, guidance_scale: float = 7.0, negative_prompt: Optional[Union[str, List[str]]] = None, negative_prompt_2: Optional[Union[str, List[str]]] = None, negative_prompt_3: Optional[Union[str, List[str]]] = None, num_images_per_prompt: Optional[int] = 1, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, latents: Optional[torch.FloatTensor] = None, prompt_embeds: Optional[torch.FloatTensor] = None, negative_prompt_embeds: Optional[torch.FloatTensor] = None, pooled_prompt_embeds: Optional[torch.FloatTensor] = None, negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, output_type: Optional[str] = "pil", ip_adapter_image: Optional[PipelineImageInput] = None, ip_adapter_image_embeds: Optional[torch.Tensor] = None, return_dict: bool = True, joint_attention_kwargs: Optional[Dict[str, Any]] = None, clip_skip: Optional[int] = None, callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, callback_on_step_end_tensor_inputs: List[str] = ["latents"], max_sequence_length: int = 256, mu: Optional[float] = None, scale: float = 1.0, offset: float = 0.08, ): r""" Function invoked when calling the pipeline for generation. Args: prompt (`str` or `List[str]`, *optional*): The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. instead. prompt_2 (`str` or `List[str]`, *optional*): The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is will be used instead prompt_3 (`str` or `List[str]`, *optional*): The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is will be used instead height (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): The height in pixels of the generated image. This is set to 1024 by default for the best results. width (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): The width in pixels of the generated image. This is set to 1024 by default for the best results. num_inference_steps (`int`, *optional*, defaults to 50): The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference. sigmas (`List[float]`, *optional*): Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed will be used. guidance_scale (`float`, *optional*, defaults to 7.0): Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, usually at the expense of lower image quality. negative_prompt (`str` or `List[str]`, *optional*): The prompt or prompts not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). negative_prompt_2 (`str` or `List[str]`, *optional*): The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `negative_prompt` is used instead negative_prompt_3 (`str` or `List[str]`, *optional*): The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `negative_prompt` is used instead num_images_per_prompt (`int`, *optional*, defaults to 1): The number of images to generate per prompt. generator (`torch.Generator` or `List[torch.Generator]`, *optional*): One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation deterministic. latents (`torch.FloatTensor`, *optional*): Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image generation. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor will ge generated by sampling using the supplied random `generator`. prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. negative_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. pooled_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, pooled text embeddings will be generated from `prompt` input argument. negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` input argument. ip_adapter_image (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. ip_adapter_image_embeds (`torch.Tensor`, *optional*): Pre-generated image embeddings for IP-Adapter. Should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. output_type (`str`, *optional*, defaults to `"pil"`): The output format of the generate image. Choose between [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] instead of a plain tuple. joint_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). callback_on_step_end (`Callable`, *optional*): A function that calls at the end of each denoising steps during the inference. The function is called with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by `callback_on_step_end_tensor_inputs`. callback_on_step_end_tensor_inputs (`List`, *optional*): The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the `._callback_tensor_inputs` attribute of your pipeline class. max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. mu (`float`, *optional*): `mu` value used for `dynamic_shifting`. Examples: Returns: [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images. """ height = height or self.default_sample_size * self.vae_scale_factor width = width or self.default_sample_size * self.vae_scale_factor # 1. Check inputs. Raise error if not correct self.check_inputs( prompt, prompt_2, prompt_3, height, width, strength, negative_prompt=negative_prompt, negative_prompt_2=negative_prompt_2, negative_prompt_3=negative_prompt_3, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds, negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, max_sequence_length=max_sequence_length, ) self._guidance_scale = guidance_scale self._clip_skip = clip_skip self._joint_attention_kwargs = joint_attention_kwargs self._interrupt = False # 2. Define call parameters if prompt is not None and isinstance(prompt, str): batch_size = 1 elif prompt is not None and isinstance(prompt, list): batch_size = len(prompt) else: batch_size = prompt_embeds.shape[0] device = self._execution_device lora_scale = ( self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None ) ( pos_prompt_embeds, _, pooled_prompt_embeds, _, ) = self.encode_prompt( prompt=prompt, prompt_2=prompt, prompt_3=prompt, do_classifier_free_guidance=False, ) ( neg_prompt_embeds, _, neg_pooled_prompt_embeds, _, ) = self.encode_prompt( prompt=negative_prompt, prompt_2=negative_prompt, prompt_3=negative_prompt, do_classifier_free_guidance=False, padding=False ) prompt_embeds = torch.cat([pos_prompt_embeds, neg_prompt_embeds], dim=1) neg_len = neg_prompt_embeds.shape[1] pos_len = prompt_embeds.shape[1] img_len = (height // 8 // self.transformer.config.patch_size) * (width // 8 //self.transformer.config.patch_size) prompt_embeds = torch.cat([prompt_embeds, neg_prompt_embeds], dim=1) attn_mask = torch.zeros((1, img_len + prompt_embeds.shape[1], img_len + prompt_embeds.shape[1] + neg_len)) attn_mask[:,-neg_len-pos_len:,-neg_len:] = -torch.inf #prompts cannot see -neg attn_mask[:,:-neg_len,-2*neg_len:-neg_len] = -torch.inf # image and positive prompt cannot see neg attn_mask[:,-neg_len:,img_len:img_len+pos_len] = -torch.inf # neg cannot see positive prompt attn_mask[:,:img_len,-neg_len:] -= offset # 0.08 image seeing less -neg attn_mask = attn_mask.to(device=device, dtype=prompt_embeds.dtype) processors_backup = [] self.maps = [] self.images = [] for block in self.transformer.transformer_blocks: processors_backup.append(block.attn.processor) block.attn.processor = JointAttnProcessor2_0(scale=scale, attn_mask=attn_mask, neg_prompt_length=neg_len, maps=self.maps) if self.do_classifier_free_guidance: prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) # 3. Preprocess image image = self.image_processor.preprocess(image, height=height, width=width) # 4. Prepare timesteps scheduler_kwargs = {} if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None: image_seq_len = (int(height) // self.vae_scale_factor // self.transformer.config.patch_size) * ( int(width) // self.vae_scale_factor // self.transformer.config.patch_size ) mu = calculate_shift( image_seq_len, self.scheduler.config.get("base_image_seq_len", 256), self.scheduler.config.get("max_image_seq_len", 4096), self.scheduler.config.get("base_shift", 0.5), self.scheduler.config.get("max_shift", 1.16), ) scheduler_kwargs["mu"] = mu elif mu is not None: scheduler_kwargs["mu"] = mu timesteps, num_inference_steps = retrieve_timesteps( self.scheduler, num_inference_steps, device, sigmas=sigmas, **scheduler_kwargs ) timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) # 5. Prepare latent variables if latents is None: latents = self.prepare_latents( image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator, ) # 6. Prepare image embeddings if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None: ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( ip_adapter_image, ip_adapter_image_embeds, device, batch_size * num_images_per_prompt, self.do_classifier_free_guidance, ) if self.joint_attention_kwargs is None: self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds} else: self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds) # 7. Denoising loop num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) self._num_timesteps = len(timesteps) with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): if self.interrupt: continue # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents # broadcast to batch dimension in a way that's compatible with ONNX/Core ML timestep = t.expand(latent_model_input.shape[0]) noise_pred = self.transformer( hidden_states=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds, pooled_projections=pooled_prompt_embeds, joint_attention_kwargs=self.joint_attention_kwargs, return_dict=False, )[0] # perform guidance if self.do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) # compute the previous noisy sample x_t -> x_t-1 latents_dtype = latents.dtype latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] if latents.dtype != latents_dtype: if torch.backends.mps.is_available(): # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 latents = latents.to(latents_dtype) if callback_on_step_end is not None: callback_kwargs = {} for k in callback_on_step_end_tensor_inputs: callback_kwargs[k] = locals()[k] callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) latents = callback_outputs.pop("latents", latents) prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) negative_pooled_prompt_embeds = callback_outputs.pop( "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds ) # 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 XLA_AVAILABLE: xm.mark_step() if output_type == "latent": image = latents else: latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor image = self.vae.decode(latents, return_dict=False)[0] image = self.image_processor.postprocess(image, output_type=output_type) # Offload all models self.maybe_free_model_hooks() if not return_dict: return (image,) for i, blocks in enumerate(self.transformer.transformer_blocks): blocks.attn.processor = processors_backup[i] return StableDiffusion3PipelineOutput(images=image)