Update app.py

app.py

@@ -11,7 +11,6 @@ import random
 import gc
 from torchao.quantization import quantize_
 from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, Int8WeightOnlyConfig
-import aoti
 
 # ------------------------
 # إعدادات النموذج
@@ -26,51 +25,75 @@ MAX_SEED = np.iinfo(np.int32).max
 FIXED_FPS = 16
 MIN_FRAMES_MODEL = 8
 MAX_FRAMES_MODEL = 480
-MIN_DURATION = round(MIN_FRAMES_MODEL/FIXED_FPS, 1)
-MAX_DURATION = round(MAX_FRAMES_MODEL/FIXED_FPS, 1)
+MIN_DURATION = round(MIN_FRAMES_MODEL / FIXED_FPS, 1)
+MAX_DURATION = round(MAX_FRAMES_MODEL / FIXED_FPS, 1)
 
 # ------------------------
 # تحميل النموذج
 # ------------------------
-pipe = WanImageToVideoPipeline.from_pretrained(MODEL_ID,
-    transformer=WanTransformer3DModel.from_pretrained('cbensimon/Wan2.2-I2V-A14B-bf16-Diffusers',
+print("🔹 Loading model... Please wait, 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='cuda'),
-    transformer_2=WanTransformer3DModel.from_pretrained('cbensimon/Wan2.2-I2V-A14B-bf16-Diffusers',
+        torch_dtype=torch.float16,
+        device_map='cuda'
+    ),
+    transformer_2=WanTransformer3DModel.from_pretrained(
+        'cbensimon/Wan2.2-I2V-A14B-bf16-Diffusers',
         subfolder='transformer_2',
-        torch_dtype=torch.bfloat16,
-        device_map='cuda'),
-    torch_dtype=torch.bfloat16,
+        torch_dtype=torch.float16,
+        device_map='cuda'
+    ),
+    torch_dtype=torch.float16
 ).to('cuda')
 
 pipe.load_lora_weights(
-    "Kijai/WanVideo_comfy", 
-    weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors", 
+    "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
+    "Kijai/WanVideo_comfy",
+    weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
+    adapter_name="lightx2v_2",
+    load_into_transformer_2=True
 )
-pipe.set_adapters(["lightx2v", "lightx2v_2"], adapter_weights=[1., 1.])
-pipe.fuse_lora(adapter_names=["lightx2v"], lora_scale=3., components=["transformer"])
-pipe.fuse_lora(adapter_names=["lightx2v_2"], lora_scale=1., components=["transformer_2"])
-pipe.unload_lora_weights()
+pipe.set_adapters(["lightx2v", "lightx2v_2"], adapter_weights=[1.0, 1.0])
+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"])
 
-quantize_(pipe.text_encoder, Int8WeightOnlyConfig())
-quantize_(pipe.transformer, Float8DynamicActivationFloat8WeightConfig())
-quantize_(pipe.transformer_2, Float8DynamicActivationFloat8WeightConfig())
-aoti.aoti_blocks_load(pipe.transformer, 'zerogpu-aoti/Wan2', variant='fp8da')
-aoti.aoti_blocks_load(pipe.transformer_2, 'zerogpu-aoti/Wan2', variant='fp8da')
+# لا نقوم بفك تحميل الـ LoRA بعد الدمج
 
 # ------------------------
-# الموجهات
+# كوانتاز اختياري (تسريع وتحسين الذاكرة)
 # ------------------------
-default_prompt_i2v = "ultra realistic cinematic footage, perfectly preserved facial identity and body structure across all frames..."
-default_negative_prompt = "low quality, low resolution, poor lighting, flickering, artifacted highlights, etc."
+if torch.cuda.is_available():
+    try:
+        quantize_(pipe.text_encoder, Int8WeightOnlyConfig())
+        quantize_(pipe.transformer, Float8DynamicActivationFloat8WeightConfig())
+        quantize_(pipe.transformer_2, Float8DynamicActivationFloat8WeightConfig())
+        print("✅ Quantization applied successfully.")
+    except Exception as e:
+        print(f"⚠️ Quantization skipped due to: {e}")
+
+# ------------------------
+# الموجهات الافتراضية
+# ------------------------
+default_prompt_i2v = (
+    "ultra realistic cinematic footage, perfectly preserved facial identity and body structure "
+    "across all frames, stable anatomy and consistent body proportions, realistic muscle definition, "
+    "natural motion flow and breathing dynamics, seamless motion continuity, photorealistic clothing "
+    "preservation with accurate fabric movement and lighting response, consistent outfit color and texture, "
+    "high-fidelity skin texture, detailed lighting and shadows"
+)
+
+default_negative_prompt = (
+    "low quality, low resolution, poor lighting, underexposed, overexposed, noise, flickering, artifacts, "
+    "stutter, inconsistent motion, broken motion, distorted face, changing face, unnatural anatomy"
+)
 
 # ------------------------
 # أدوات الصورة والفيديو
@@ -79,30 +102,31 @@ def resize_image(image: Image.Image) -> Image.Image:
     width, height = image.size
     if width == height:
         return image.resize((SQUARE_DIM, SQUARE_DIM), Image.LANCZOS)
+
     aspect_ratio = width / height
     MAX_AR = MAX_DIM / MIN_DIM
     MIN_AR = MIN_DIM / MAX_DIM
-    image_to_resize = image
+
     if aspect_ratio > MAX_AR:
-        target_w, target_h = MAX_DIM, MIN_DIM
         crop_width = int(round(height * MAX_AR))
         left = (width - crop_width) // 2
-        image_to_resize = image.crop((left, 0, left + crop_width, height))
+        image = image.crop((left, 0, left + crop_width, height))
     elif aspect_ratio < MIN_AR:
-        target_w, target_h = MIN_DIM, MAX_DIM
         crop_height = int(round(width / MIN_AR))
         top = (height - crop_height) // 2
-        image_to_resize = image.crop((0, top, width, top + crop_height))
+        image = image.crop((0, top, width, top + crop_height))
+
+    if width > height:
+        target_w = MAX_DIM
+        target_h = int(round(target_w / aspect_ratio))
     else:
-        if width > height:
-            target_w = MAX_DIM
-            target_h = int(round(target_w / aspect_ratio))
-        else:
-            target_h = MAX_DIM
-            target_w = int(round(target_h * aspect_ratio))
+        target_h = MAX_DIM
+        target_w = int(round(target_h * aspect_ratio))
+
     final_w = max(MIN_DIM, min(MAX_DIM, round(target_w / MULTIPLE_OF) * MULTIPLE_OF))
     final_h = max(MIN_DIM, min(MAX_DIM, round(target_h / MULTIPLE_OF) * MULTIPLE_OF))
-    return image_to_resize.resize((final_w, final_h), Image.LANCZOS)
+    return image.resize((final_w, final_h), Image.LANCZOS)
+
 
 def get_num_frames(duration_seconds: float):
     return 1 + int(np.clip(int(round(duration_seconds * FIXED_FPS)), MIN_FRAMES_MODEL, MAX_FRAMES_MODEL))
@@ -111,16 +135,27 @@ def get_num_frames(duration_seconds: float):
 # عملية التوليد
 # ------------------------
 @spaces.GPU()
-def generate_video(input_image, prompt, steps=4, negative_prompt=default_negative_prompt,
-                   duration_seconds=3.5, guidance_scale=1, guidance_scale_2=1,
-                   seed=42, randomize_seed=False, progress=gr.Progress(track_tqdm=True)):
+def generate_video(
+    input_image,
+    prompt,
+    steps=4,
+    negative_prompt=default_negative_prompt,
+    duration_seconds=3.5,
+    guidance_scale=1.0,
+    guidance_scale_2=1.0,
+    seed=42,
+    randomize_seed=False,
+    progress=gr.Progress(track_tqdm=True)
+):
     if input_image is None:
-        raise gr.Error("Please upload an input image.")
+        raise gr.Error("⚠️ Please upload an input image first.")
+
     num_frames = get_num_frames(duration_seconds)
     current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
     resized_image = resize_image(input_image)
+
     with progress.tqdm(total=100) as pbar:
-        pbar.set_description("Generating video...")
+        pbar.set_description("🎬 Generating video...")
         output_frames_list = pipe(
             image=resized_image,
             prompt=prompt,
@@ -133,9 +168,13 @@ def generate_video(input_image, prompt, steps=4, negative_prompt=default_negativ
             num_inference_steps=int(steps),
             generator=torch.Generator(device="cuda").manual_seed(current_seed),
         ).frames[0]
+
     with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
         video_path = tmpfile.name
+
     export_to_video(output_frames_list, video_path, fps=FIXED_FPS)
+    torch.cuda.empty_cache()
+    gc.collect()
     return video_path, current_seed
 
 # ------------------------
@@ -144,34 +183,36 @@ def generate_video(input_image, prompt, steps=4, negative_prompt=default_negativ
 with gr.Blocks(theme=gr.themes.Soft(primary_hue="violet")) as demo:
     gr.HTML("""
     <div style="text-align:center; padding:20px;">
-        <h1 style="font-size: 2em;">Wan 2.2 Lightning Studio – AI Cinematic Video Generator - dream2589632147</h1>
-        <p style="opacity:0.8;">Powered by dream2589632147 ⚡</p>
+        <h1 style="font-size: 2em;">Wan 2.2 Lightning Studio – AI Cinematic Video Generator</h1>
+        <p style="opacity:0.8;">⚡ Powered by dream2589632147</p>
     </div>
     """)
-    
+
     with gr.Row():
         with gr.Column(scale=1):
             input_image = gr.Image(label="🎞️ Input Image", type="pil")
             prompt = gr.Textbox(label="✨ Positive Prompt", value=default_prompt_i2v, lines=3)
             negative_prompt = gr.Textbox(label="🚫 Negative Prompt", value=default_negative_prompt, lines=3)
-            duration = gr.Slider(MIN_DURATION, MAX_DURATION, value=3.5, step=0.1, label="Duration (seconds)")
+            duration = gr.Slider(MIN_DURATION, MAX_DURATION, value=3.5, step=0.1, label="🎬 Duration (seconds)")
             with gr.Accordion("⚙️ Advanced Settings", open=False):
                 steps = gr.Slider(1, 30, value=6, step=1, label="Inference Steps")
-                guidance_scale = gr.Slider(0.0, 10.0, value=1, step=0.5, label="Guidance Scale 1")
-                guidance_scale_2 = gr.Slider(0.0, 10.0, value=1, step=0.5, label="Guidance Scale 2")
+                guidance_scale = gr.Slider(0.0, 10.0, value=1.0, step=0.5, label="Guidance Scale 1")
+                guidance_scale_2 = gr.Slider(0.0, 10.0, value=1.0, step=0.5, label="Guidance Scale 2")
                 seed = gr.Slider(0, MAX_SEED, value=42, step=1, label="Seed")
                 randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
             generate_btn = gr.Button("🚀 Generate Cinematic Video", variant="primary")
 
         with gr.Column(scale=1):
-            progress_text = gr.Textbox(label="Progress", interactive=False)
             video_output = gr.Video(label="🎬 Generated Video Preview", autoplay=True)
+            seed_output = gr.Textbox(label="🎲 Seed Used", interactive=False)
             download_btn = gr.File(label="⬇️ Download MP4")
 
-    generate_btn.click(fn=generate_video,
-                       inputs=[input_image, prompt, steps, negative_prompt, duration,
-                               guidance_scale, guidance_scale_2, seed, randomize_seed],
-                       outputs=[video_output, seed])
+    generate_btn.click(
+        fn=generate_video,
+        inputs=[input_image, prompt, steps, negative_prompt, duration,
+                guidance_scale, guidance_scale_2, seed, randomize_seed],
+        outputs=[video_output, seed_output]
+    )
 
     # زر تبديل الوضع الليلي/النهاري
     gr.HTML("""