Update app.py

app.py

@@ -135,8 +135,6 @@ def interpolate_bits(frames_np, multiplier=2, scale=1.0):
     
     # Handle input shape
     if isinstance(frames_np, list):
-        # Convert list of arrays to one big array for easier shape handling if needed, 
-        # but here we just grab dims from first frame
         T = len(frames_np)
         H, W, C = frames_np[0].shape
     else:
@@ -144,7 +142,6 @@ def interpolate_bits(frames_np, multiplier=2, scale=1.0):
 
     # 1. No Interpolation Case
     if multiplier < 2:
-        # Just convert 4D array to list of 3D arrays
         if isinstance(frames_np, np.ndarray):
             return list(frames_np)
         return frames_np
@@ -159,19 +156,14 @@ def interpolate_bits(frames_np, multiplier=2, scale=1.0):
 
     # Helper: Numpy (H, W, C) Float -> Tensor (1, C, H, W) Half
     def to_tensor(frame_np):
-        # frame_np is float32 0-1
         t = torch.from_numpy(frame_np).to(device)
-        # HWC -> CHW
         t = t.permute(2, 0, 1).unsqueeze(0)
         return F.pad(t, padding).half()
 
     # Helper: Tensor (1, C, H, W) Half -> Numpy (H, W, C) Float
     def from_tensor(tensor):
-        # Crop padding
         t = tensor[0, :, :H, :W]
-        # CHW -> HWC
         t = t.permute(1, 2, 0)
-        # Keep as float32, range 0-1
         return t.float().cpu().numpy()
 
     def make_inference(I0, I1, n):
@@ -194,25 +186,17 @@ def interpolate_bits(frames_np, multiplier=2, scale=1.0):
     output_frames =[]
 
     # Process Frames
-    # Load first frame into GPU
     I1 = to_tensor(frames_np[0])
-
     total_steps = T - 1
 
     with tqdm(total=total_steps, desc="Interpolating", unit="frame") as pbar:
-    
         for i in range(total_steps):
             I0 = I1
-            # Add original frame to output
             output_frames.append(from_tensor(I0))
     
-            # Load next frame
             I1 = to_tensor(frames_np[i+1])
-    
-            # Generate intermediate frames
             mid_tensors = make_inference(I0, I1, n_interp)
     
-            # Append intermediate frames
             for mid in mid_tensors:
                 output_frames.append(from_tensor(mid))
 
@@ -220,7 +204,6 @@ def interpolate_bits(frames_np, multiplier=2, scale=1.0):
                 pbar.update(50)
         pbar.update(total_steps % 50)
         
-        # Add the very last frame
         output_frames.append(from_tensor(I1))
     
     # Cleanup
@@ -264,21 +247,10 @@ pipe = WanImageToVideoPipeline.from_pretrained(
 ).to('cuda')
 original_scheduler = copy.deepcopy(pipe.scheduler)
 
-# if os.path.exists(CACHE_DIR):
-#     shutil.rmtree(CACHE_DIR)
-#     print("Deleted Hugging Face cache.")
-# else:
-#     print("No hub cache found.")
-
 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')
-
-# pipe.vae.enable_slicing()
-# pipe.vae.enable_tiling()
 
 default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
 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"
@@ -382,7 +354,6 @@ def get_inference_duration(
     return total_time
     
 
-
 @spaces.GPU(duration=get_inference_duration, size='xlarge')
 def run_inference(
     resized_image,
@@ -431,7 +402,7 @@ def run_inference(
         output_type="np" 
     )
     
-    raw_frames_np = result.frames[0]  # Returns (T, H, W, C) float32
+    raw_frames_np = result.frames[0]
     pipe.scheduler = original_scheduler
 
     frame_factor = frame_multiplier // FIXED_FPS
@@ -568,17 +539,55 @@ with gr.Blocks(theme=gr.themes.Soft(), css=CSS, delete_cache=(3600, 3700)) as de
             generate_button = gr.Button("Generate Video", variant="primary")
 
         with gr.Column():
-            # ASSIGNED elem_id="generated-video" so JS can find it
             video_output = gr.Video(label="Generated Video", autoplay=True, sources=["upload"], show_download_button=True, show_share_button=True, interactive=False, elem_id="generated-video")
             
-            # --- Frame Grabbing UI ---
             with gr.Row():
                 grab_frame_btn = gr.Button("📸 Use Current Frame as Input", variant="secondary")
                 timestamp_box = gr.Number(value=0, label="Timestamp", visible=True, elem_id="hidden-timestamp")
-            # -------------------------
             
             file_output = gr.File(label="Download Video")
 
-    ui_inputs =[
-        input_image_component, last_image_component, prompt_input, steps_slider,
-        neg
+    ui_inputs = [
+        input_image_component, 
+        last_image_component, 
+        prompt_input, 
+        steps_slider,
+        negative_prompt_input, 
+        duration_seconds_input,
+        guidance_scale_input, 
+        guidance_scale_2_input, 
+        seed_input, 
+        randomize_seed_checkbox,
+        quality_slider, 
+        scheduler_dropdown, 
+        flow_shift_slider, 
+        frame_multi,
+        play_result_video, 
+        safe_mode_checkbox
+    ]
+    
+    generate_button.click(
+        fn=generate_video, 
+        inputs=ui_inputs, 
+        outputs=[video_output, file_output, seed_input]
+    )
+    
+    grab_frame_btn.click(
+        fn=None,
+        inputs=None,
+        outputs=[timestamp_box],
+        js=get_timestamp_js
+    )
+    
+    timestamp_box.change(
+        fn=extract_frame,
+        inputs=[video_output, timestamp_box],
+        outputs=[input_image_component]
+    )
+
+if __name__ == "__main__":
+    demo.queue().launch(
+        mcp_server=True,
+        ssr_mode=False,
+        show_error=True,
+    )