Update app_wan.py

app_wan.py

@@ -1,268 +1,111 @@
-# app_wa
-import os
-# PyTorch 2.8 (temporary hack)
-os.system('pip install --upgrade --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu126 "torch<2.9" spaces')
+# app_wan.py (trechos essenciais)
 
-# --- 1. Model Download and Setup (Diffusers Backend) ---
+import os
+# (mantenha o hack de instalação se necessário)
+#os.system('pip install --upgrade --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu126 "torch<2.9" spaces')
 
-import torch
-from diffusers import FlowMatchEulerDiscreteScheduler
-from diffusers.pipelines.wan.pipeline_wan_i2v import WanImageToVideoPipeline
-from diffusers.models.transformers.transformer_wan import WanTransformer3DModel
-from diffusers.utils.export_utils import export_to_video
 import gradio as gr
-import tempfile
-import numpy as np
 from PIL import Image
-import random
-import gc
-from gradio_client import Client, handle_file # Import for API call
+import numpy as np
+import tempfile
 
-# --- Constants and Model Loading ---
+# === mantém as constantes de UI exatamente como no arquivo atual ===
 MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers"
-
-# --- NEW: Flexible Dimension Constants ---
 MAX_DIMENSION = 832
 MIN_DIMENSION = 480
 DIMENSION_MULTIPLE = 16
 SQUARE_SIZE = 480
-
 MAX_SEED = np.iinfo(np.int32).max
-
 FIXED_FPS = 16
 MIN_FRAMES_MODEL = 8
 MAX_FRAMES_MODEL = 81
+MIN_DURATION = round(MIN_FRAMES_MODEL / FIXED_FPS, 1)
+MAX_DURATION = round(MAX_FRAMES_MODEL / FIXED_FPS, 1)
+default_negative_prompt = "..."  # mesmo valor do arquivo original
 
-MIN_DURATION = round(MIN_FRAMES_MODEL/FIXED_FPS, 1)
-MAX_DURATION = round(MAX_FRAMES_MODEL/FIXED_FPS, 1)
-
-default_negative_prompt = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走,过曝，"
-
-print("Loading models into memory. This may take a few minutes...")
-
-pipe    = WanImageToVideoPipeline.from_pretrained(
-    MODEL_ID,
-    transformer=WanTransformer3DModel.from_pretrained('cbensimon/Wan2.2-I2V-A14B-bf16-Diffusers',
-        subfolder='transformer',
-        torch_dtype=torch.bfloat16,
-        device_map='auto',
-    ),
-    transformer_2=WanTransformer3DModel.from_pretrained('cbensimon/Wan2.2-I2V-A14B-bf16-Diffusers',
-        subfolder='transformer_2',
-        torch_dtype=torch.bfloat16,
-        device_map='auto',
-    ),
-    torch_dtype=torch.bfloat16,
-)
-pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(pipe.scheduler.config, shift=32.0)
-
-
-
-
-
-
-
-# ====================================================================================
-# A fusão do LoRA "Lightning" é ESSENCIAL para a geração em 8 passos.
-# Trazemos essa lógica para cá, mantendo a otimização completa desativada.
-# ====================================================================================
-
-print("Applying 8-step Lightning LoRA...")
-try:
-    # Carrega os pesos do LoRA para os dois transformadores
-    pipe.load_lora_weights(
-        "Kijai/WanVideo_comfy",
-        weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
-        adapter_name="lightx2v"
-    )
-    kwargs_lora = {"load_into_transformer_2": True}
-    pipe.load_lora_weights(
-        "Kijai/WanVideo_comfy",
-        weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
-        adapter_name="lightx2v_2", **kwargs_lora
-    )
-    
-    # Define como os adaptadores LoRA serão combinados
-    pipe.set_adapters(["lightx2v", "lightx2v_2"], adapter_weights=[1.0, 1.0])
-    
-    # Funde os LoRAs diretamente nos pesos do modelo para acelerar a inferência.
-    print("Fusing LoRA weights into the main model...")
-    pipe.fuse_lora(adapter_names=["lightx2v"], lora_scale=3.0, components=["transformer"])
-    pipe.fuse_lora(adapter_names=["lightx2v_2"], lora_scale=1.0, components=["transformer_2"])
-    
-    # Descarrega os pesos LoRA da memória, pois eles já foram incorporados.
-    pipe.unload_lora_weights()
-    
-    print("Lightning LoRA successfully fused. Model is ready for fast 8-step generation.")
-
-except Exception as e:
-    print(f"AVISO: Falha ao carregar ou fundir o LoRA. A geração pode ser lenta ou de baixa qualidade. Erro: {e}")
-
-print("All models loaded. Gradio app is ready.")
-
+# === novo: importar e instanciar o serviço ===
+from aduc_framework.managers.wan_manager import WanManager
+wan_manager = WanManager()
 
-# --- 2. Image Processing and Application Logic ---
-def generate_end_frame(start_img, gen_prompt, progress=gr.Progress(track_tqdm=True)):
-    """Calls an external Gradio API to generate an image."""
-    if start_img is None:
-        raise gr.Error("Please provide a Start Frame first.")
+# (opcional) manter generate_end_frame e switch_to_upload_tab como estão
+# def generate_end_frame(...): ...    # inalterado
+# def switch_to_upload_tab(): ...     # inalterado
 
-    hf_token = os.getenv("HF_TOKEN")
-    if not hf_token:
-        raise gr.Error("HF_TOKEN not found in environment variables. Please set it in your Space secrets.")
-
-    with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmpfile:
-        start_img.save(tmpfile.name)
-        tmp_path = tmpfile.name
-
-    progress(0.1, desc="Connecting to image generation API...")
-    client = Client("multimodalart/nano-banana")
-    
-    progress(0.5, desc=f"Generating with prompt: '{gen_prompt}'...")
-    try:
-        result = client.predict(
-            prompt=gen_prompt,
-            images=[
-                {"image": handle_file(tmp_path)}
-            ],
-            manual_token=hf_token,
-            api_name="/unified_image_generator"
-        )
-    finally:
-        os.remove(tmp_path)
-
-    progress(1.0, desc="Done!")
-    print(result)
-    return result
-
-def switch_to_upload_tab():
-    """Returns a gr.Tabs update to switch to the first tab."""
-    return gr.Tabs(selected="upload_tab")
-
-
-def process_image_for_video(image: Image.Image) -> Image.Image:
-    """
-    Resizes an image based on the following rules for video generation.
-    """
-    width, height = image.size
-    if width == height:
-        return image.resize((SQUARE_SIZE, SQUARE_SIZE), Image.Resampling.LANCZOS)
-    aspect_ratio = width / height
-    new_width, new_height = width, height
-    if new_width > MAX_DIMENSION or new_height > MAX_DIMENSION:
-        if aspect_ratio > 1: scale = MAX_DIMENSION / new_width
-        else: scale = MAX_DIMENSION / new_height
-        new_width *= scale; new_height *= scale
-    if new_width < MIN_DIMENSION or new_height < MIN_DIMENSION:
-        if aspect_ratio > 1: scale = MIN_DIMENSION / new_height
-        else: scale = MIN_DIMENSION / new_width
-        new_width *= scale; new_height *= scale
-    final_width = int(round(new_width / DIMENSION_MULTIPLE) * DIMENSION_MULTIPLE)
-    final_height = int(round(new_height / DIMENSION_MULTIPLE) * DIMENSION_MULTIPLE)
-    final_width = max(final_width, MIN_DIMENSION if aspect_ratio < 1 else SQUARE_SIZE)
-    final_height = max(final_height, MIN_DIMENSION if aspect_ratio > 1 else SQUARE_SIZE)
-    return image.resize((final_width, final_height), Image.Resampling.LANCZOS)
-
-def resize_and_crop_to_match(target_image, reference_image):
-    """Resizes and center-crops the target image to match the reference image's dimensions."""
-    ref_width, ref_height = reference_image.size
-    target_width, target_height = target_image.size
-    scale = max(ref_width / target_width, ref_height / target_height)
-    new_width, new_height = int(target_width * scale), int(target_height * scale)
-    resized = target_image.resize((new_width, new_height), Image.Resampling.LANCZOS)
-    left, top = (new_width - ref_width) // 2, (new_height - ref_height) // 2
-    return resized.crop((left, top, left + ref_width, top + ref_height))
-
-
-def generate_video(
+# Wrapper fino: a UI constrói images_condition_items e delega ao serviço
+def ui_generate_video(
     start_image_pil,
+    start_frame_text,
+    start_peso,
     end_image_pil,
+    end_frame_text,
+    end_peso,
     prompt,
     negative_prompt=default_negative_prompt,
     duration_seconds=2.1,
     steps=8,
-    guidance_scale=1,
-    guidance_scale_2=1,
+    guidance_scale=1.0,
+    guidance_scale_2=1.0,
     seed=42,
     randomize_seed=False,
     progress=gr.Progress(track_tqdm=True)
 ):
-    """
-    Generates a video by interpolating between a start and end image, guided by a text prompt.
-    """
-    if start_image_pil is None or end_image_pil is None:
-        raise gr.Error("Please upload both a start and an end image.")
-
-    progress(0.1, desc="Preprocessing images...")
-    processed_start_image = process_image_for_video(start_image_pil)
-    processed_end_image = resize_and_crop_to_match(end_image_pil, processed_start_image)
-    target_height, target_width = processed_start_image.height, processed_start_image.width
-
-    current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
-    num_frames = np.clip(int(round(duration_seconds * FIXED_FPS)), MIN_FRAMES_MODEL, MAX_FRAMES_MODEL)
-
-    progress(0.2, desc=f"Generating {num_frames} frames at {target_width}x{target_height} (seed: {current_seed})...")
-
-    # CORREÇÃO FINAL: O gerador é criado na CPU (padrão) para evitar o erro de dispositivo 'meta'.
-    # A pipeline cuidará de mover os latentes para a GPU.
-    generator = torch.Generator().manual_seed(current_seed)
-    
-    output_frames_list = pipe(
-        image=processed_start_image,
-        last_image=processed_end_image,
+    # Conversão robusta básica (UI “burra”, sem semântica adicional)
+    def to_int_safe(v, default=0):
+        try:
+            return int(v)
+        except:
+            return default
+    def to_float_safe(v, default=1.0):
+        try:
+            return float(v)
+        except:
+            return default
+
+    images_condition_items = [
+        [start_image_pil, to_int_safe(start_frame_text, 0), to_float_safe(start_peso, 1.0)],
+        [end_image_pil,   to_int_safe(end_frame_text,   0), to_float_safe(end_peso,   1.0)],
+    ]
+    return wan_manager.generate_video_from_conditions(
+        images_condition_items=images_condition_items,
         prompt=prompt,
         negative_prompt=negative_prompt,
-        height=target_height,
-        width=target_width,
-        num_frames=num_frames,
-        guidance_scale=float(guidance_scale),
-        guidance_scale_2=float(guidance_scale_2),
-        num_inference_steps=int(steps),
-        generator=generator,
-    ).frames[0]
-
-    progress(0.9, desc="Encoding and saving video...")
-    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
-        video_path = tmpfile.name
-    export_to_video(output_frames_list, video_path, fps=FIXED_FPS)
-
-    progress(1.0, desc="Done!")
-    return video_path, current_seed
-
-
-# --- 3. Gradio User Interface ---
+        duration_seconds=duration_seconds,
+        steps=steps,
+        guidance_scale=guidance_scale,
+        guidance_scale_2=guidance_scale_2,
+        seed=seed,
+        randomize_seed=randomize_seed,
+    )
 
-css = '''
-.fillable{max-width: 1100px !important}
-.dark .progress-text {color: white}
-#general_items{margin-top: 2em}
-#group_all{overflow:visible}
-#group_all .styler{overflow:visible}
-#group_tabs .tabitem{padding: 0}
-.tab-wrapper{margin-top: -33px;z-index: 999;position: absolute;width: 100%;background-color: var(--block-background-fill);padding: 0;}
-#component-9-button{width: 50%;justify-content: center}
-#component-11-button{width: 50%;justify-content: center}
-#or_item{text-align: center; padding-top: 1em; padding-bottom: 1em; font-size: 1.1em;margin-left: .5em;margin-right: .5em;width: calc(100% - 1em)}
-#fivesec{margin-top: 5em;margin-left: .5em;margin-right: .5em;width: calc(100% - 1em)}
-'''
+# === UI Gradio (adicionar frame/peso abaixo de cada imagem) ===
+css = ''' ... (mantém igual) ... '''
 with gr.Blocks(theme=gr.themes.Citrus(), css=css) as app:
     gr.Markdown("# Wan 2.2 Aduca-sdr")
-    
+
     with gr.Row(elem_id="general_items"):
         with gr.Column():
             with gr.Group(elem_id="group_all"):
                 with gr.Row():
-                    start_image = gr.Image(type="pil", label="Start Frame", sources=["upload", "clipboard"])
+                    with gr.Column():
+                        start_image = gr.Image(type="pil", label="Start Frame", sources=["upload", "clipboard"])
+                        start_frame_tb = gr.Textbox(label="Start Frame (int)", value="0")
+                        start_peso_sl = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, value=1.0, label="Start Peso")
                     with gr.Tabs(elem_id="group_tabs") as tabs:
                         with gr.TabItem("Upload", id="upload_tab"):
-                            end_image = gr.Image(type="pil", label="End Frame", sources=["upload", "clipboard"])
+                            with gr.Column():
+                                end_image = gr.Image(type="pil", label="End Frame", sources=["upload", "clipboard"])
+                                end_frame_tb = gr.Textbox(label="End Frame (int)", value="80")
+                                end_peso_sl = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, value=1.0, label="End Peso")
                         with gr.TabItem("Generate", id="generate_tab"):
                             generate_5seconds = gr.Button("Generate scene 5 seconds in the future", elem_id="fivesec")
-                            gr.Markdown("Generate a custom end-frame with an edit model like [Nano Banana](https://huggingface.co/spaces/multimodalart/nano-banana) or [Qwen Image Edit](https://huggingface.co/spaces/multimodalart/Qwen-Image-Edit-Fast)", elem_id="or_item")
+                            gr.Markdown(
+                                "Generate a custom end-frame with an edit model like Nano Banana or Qwen Image Edit",
+                                elem_id="or_item"
+                            )
                 prompt = gr.Textbox(label="Prompt", info="Describe the transition between the two images")
 
                 with gr.Accordion("Advanced Settings", open=False):
-                    duration_seconds_input = gr.Slider(minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=2.1, label="Video Duration (seconds)", info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps.")
+                    duration_seconds_input = gr.Slider(minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=2.1, label="Video Duration (seconds)")
                     negative_prompt_input = gr.Textbox(label="Negative Prompt", value=default_negative_prompt, lines=3)
                     steps_slider = gr.Slider(minimum=1, maximum=30, step=1, value=8, label="Inference Steps")
                     guidance_scale_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1.0, label="Guidance Scale - high noise")
@@ -276,29 +119,35 @@ with gr.Blocks(theme=gr.themes.Citrus(), css=css) as app:
         with gr.Column():
             output_video = gr.Video(label="Generated Video", autoplay=True)
 
+    # Inputs/outputs atualizados para o wrapper
     ui_inputs = [
-        start_image, end_image, prompt, negative_prompt_input, duration_seconds_input,
-        steps_slider, guidance_scale_input, guidance_scale_2_input, seed_input,
-        randomize_seed_checkbox
+        start_image, start_frame_tb, start_peso_sl,
+        end_image, end_frame_tb, end_peso_sl,
+        prompt, negative_prompt_input, duration_seconds_input,
+        steps_slider, guidance_scale_input, guidance_scale_2_input,
+        seed_input, randomize_seed_checkbox
     ]
     ui_outputs = [output_video, seed_input]
 
-    generate_button.click(fn=generate_video, inputs=ui_inputs, outputs=ui_outputs)
+    generate_button.click(fn=ui_generate_video, inputs=ui_inputs, outputs=ui_outputs)
 
+    # Cadeia “5 seconds”: mantém a lógica, apenas troca a função final para ui_generate_video
     generate_5seconds.click(
-        fn=switch_to_upload_tab,
-        inputs=None,
-        outputs=[tabs]
+        fn=switch_to_upload_tab, inputs=None, outputs=[tabs]
     ).then(
-        fn=lambda img: generate_end_frame(img, "this image is a still frame from a movie. generate a new frame with what happens on this scene 5 seconds in the future"),
+        fn=lambda img: generate_end_frame(
+            img,
+            "this image is a still frame from a movie. generate a new frame with what happens on this scene 5 seconds in the future"
+        ),
         inputs=[start_image],
         outputs=[end_image]
     ).success(
-        fn=generate_video,
+        fn=ui_generate_video,
         inputs=ui_inputs,
         outputs=ui_outputs
     )
 
+    # Examples permanecem; os campos de frame/peso usam defaults
     gr.Examples(
         examples=[
             ["poli_tower.png", "tower_takes_off.png", "the man turns around"],
@@ -307,28 +156,6 @@ with gr.Blocks(theme=gr.themes.Citrus(), css=css) as app:
         ],
         inputs=[start_image, end_image, prompt],
         outputs=ui_outputs,
-        fn=generate_video,
+        fn=ui_generate_video,
         cache_examples="lazy",
     )
-
-
-    
-
-if __name__ == "__main__":
-    # Cria a pasta e as imagens de exemplo se não existirem
-    if not os.path.exists("examples"):
-        os.makedirs("examples")
-    try:
-        Image.new('RGB', (832, 480), color = (73, 109, 137)).save("examples/frame_1.png")
-        Image.new('RGB', (832, 480), color = (173, 109, 237)).save("examples/frame_2.png")
-        Image.new('RGB', (832, 480), color = (255, 255, 0)).save("examples/frame_3.png")
-    except:
-        pass # Evita que o app quebre se não tiver permissão de escrita
-
-    os.makedirs("deformes_workspace", exist_ok=True)
-    #logger.info("Aplicação Gradio pronta. Lançando interface...")
-    app.launch(
-        debug=True,
-        server_name="0.0.0.0",
-        server_port=int(os.getenv("PORT", "7860")),
-    )