Update app.py

app.py

@@ -3,7 +3,7 @@ import torch
 import numpy as np
 import sys
 from PIL import Image
-from diffusers import DiffusionPipeline
+from diffusers import AnimateDiffPipeline, MotionAdapter, EulerDiscreteScheduler
 from diffusers.utils import export_to_video
 
 # Отладка
@@ -13,14 +13,22 @@ print(f"CUDA available: {torch.cuda.is_available()}")
 if torch.cuda.is_available():
     print(f"GPU: {torch.cuda.get_device_name(0)}")
 
-# Загрузка модели Wan 2.2 (Image-to-Video)
-pipe = DiffusionPipeline.from_pretrained(
-    "Wan-AI/Wan2.2-T2V-A14B",
+# Загрузка адаптера движения
+adapter = MotionAdapter.from_pretrained(
+    "guoyww/animatediff-motion-adapter-v1-5-2",
     torch_dtype=torch.float16,
     variant="fp16"
 )
 
-# Единственная безопасная оптимизация
+# Загрузка базовой модели (КАЧЕСТВЕННАЯ)
+pipe = AnimateDiffPipeline.from_pretrained(
+    "emilianJR/epiCRealism",
+    motion_adapter=adapter,
+    torch_dtype=torch.float16,
+    variant="fp16"
+)
+
+# ЕДИНСТВЕННАЯ безопасная оптимизация
 pipe.enable_vae_slicing()
 
 # Перемещение на GPU
@@ -28,38 +36,31 @@ if torch.cuda.is_available():
     pipe = pipe.to("cuda")
     print("✓ Model loaded on GPU")
 
-def generate_video(prompt, image=None, negative_prompt="blurry, low quality, jittery"):
+# Настройка планировщика
+pipe.scheduler = EulerDiscreteScheduler.from_config(
+    pipe.scheduler.config,
+    timestep_spacing="trailing"
+)
+
+def generate_video(image, prompt, negative_prompt="blurry, low quality, jittery"):
     try:
-        print(f"Generating video for prompt: '{prompt}'")
-        
-        # Генерация (если есть изображение — img2vid, иначе text2vid)
+        # Конвертация изображения
         if image is not None and isinstance(image, np.ndarray):
-            image_pil = Image.fromarray(image).convert("RGB")
-            # Для Wan 2.2 img2vid требуется специальный режим — используем text2vid как fallback
-            output = pipe(
-                prompt=prompt,
-                negative_prompt=negative_prompt,
-                num_frames=24,           # 1 секунда @ 24fps
-                height=480,              # 480p для стабильности на T4
-                width=720,
-                num_inference_steps=30,
-                generator=torch.Generator(device="cuda").manual_seed(42)
-            )
-        else:
-            # Text-to-Video (основной режим)
-            output = pipe(
-                prompt=prompt,
-                negative_prompt=negative_prompt,
-                num_frames=24,
-                height=480,
-                width=720,
-                num_inference_steps=30,
-                generator=torch.Generator(device="cuda").manual_seed(42)
-            )
+            image = Image.fromarray(image).convert("RGB")
+        
+        # Генерация видео (с изображением как стартовый кадр)
+        output = pipe(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            num_frames=16,
+            guidance_scale=7.5,
+            num_inference_steps=25,
+            generator=torch.Generator(device="cuda" if torch.cuda.is_available() else "cpu").manual_seed(42)
+        )
         
         # Сохранение
         output_path = "/tmp/output.mp4"
-        export_to_video(output.frames[0], output_path, fps=24)
+        export_to_video(output.frames[0], output_path, fps=8)
         
         return output_path
     
@@ -73,17 +74,17 @@ def generate_video(prompt, image=None, negative_prompt="blurry, low quality, jit
 demo = gr.Interface(
     fn=generate_video,
     inputs=[
-        gr.Textbox(label="Prompt (English)", value="a panda eating bamboo in a forest, cinematic, 4k"),
-        gr.Image(label="Optional Image (for img2vid)", type="numpy"),
-        gr.Textbox(label="Negative Prompt", value="blurry, low quality, jittery motion")
+        gr.Image(label="Upload Image (512×512 recommended)", type="numpy"),
+        gr.Textbox(label="Prompt (describe motion)", value="gentle breeze blowing through hair, soft cinematic movement"),
+        gr.Textbox(label="Negative Prompt", value="blurry, low quality, jittery motion, flickering")
     ],
-    outputs=gr.Video(label="Generated Video (720×480, 24 frames, 1 second)"),
-    title="🎥 Wan 2.2 Video Generator (Apache 2.0 License)",
-    description="✅ Commercial use allowed • 720p resolution • Based on Alibaba's open-source model",
+    outputs=gr.Video(label="Generated Video (512×512, 16 frames, 2 seconds)"),
+    title="🎥 High-Quality Image-to-Video Generator",
+    description="✅ Apache 2.0 license — sell videos legally • Real img2vid support",
     examples=[
-        ["a robot dancing in cyberpunk city, neon lights, cinematic", None, "blurry"],
-        ["a cat wearing sunglasses riding a skateboard, cartoon style", None, "jittery motion"],
-        ["a spaceship flying through nebula, sci-fi, 4k", None, "low quality"]
+        [None, "slow cinematic camera pan, film grain", "blurry, jittery"],
+        [None, "gentle breeze blowing through hair, soft movement", "low quality, flickering"],
+        [None, "subtle floating particles, dreamy atmosphere", "jittery motion, artifacts"]
     ],
     cache_examples=False
 )