update

.gitignore

@@ -197,4 +197,5 @@ slurm-*.out
 .vscode
 
 data/
-tmp/
+tmp/
+.gradio/

app.py

@@ -33,6 +33,9 @@ from submodules.MoGe.moge.model import MoGeModel
 from submodules.vggt.vggt.utils.pose_enc import pose_encoding_to_extri_intri
 from submodules.vggt.vggt.models.vggt import VGGT
 
+import torch._dynamo
+torch._dynamo.config.suppress_errors = True
+
 # Parse command line arguments
 parser = argparse.ArgumentParser(description="Diffusion as Shader Web UI")
 parser.add_argument("--port", type=int, default=7860, help="Port to run the web UI on")
@@ -82,9 +85,9 @@ def load_media(media_path, max_frames=49, transform=None):
         
         # Case 1: Video longer than 6 seconds, sample first 6 seconds + 1 frame
         if duration > 6.0:
-            sampling_fps = 8  # 8 frames per second
-            frames = load_video(media_path, sampling_fps=sampling_fps, max_frames=max_frames)
-            fps = sampling_fps
+            # 使用 max_frames 参数而不是 sampling_fps
+            frames = load_video(media_path, max_frames=max_frames)
+            fps = max_frames / 6.0  # 计算等效的 fps
         # Cases 2 and 3: Video shorter than 6 seconds
         else:
             # Load all frames
@@ -195,10 +198,10 @@ def get_vggt_model():
 def process_motion_transfer(source, prompt, mt_repaint_option, mt_repaint_image):
     """Process video motion transfer task"""
     try:
-        # Save uploaded files
+        # 保存上传的文件
         input_video_path = save_uploaded_file(source)
         if input_video_path is None:
-            return None, None
+            return None, None, None
         
         print(f"DEBUG: Repaint option: {mt_repaint_option}")
         print(f"DEBUG: Repaint image: {mt_repaint_image}")
@@ -253,28 +256,20 @@ def process_motion_transfer(source, prompt, mt_repaint_option, mt_repaint_image)
             tracking_path, tracking_tensor = das.visualize_tracking_cotracker(pred_tracks, pred_visibility)
             print('Export tracking video via cotracker')
 
-        output_path = das.apply_tracking(
-            video_tensor=video_tensor,
-            fps=fps,  # 使用 load_media 返回的 fps
-            tracking_tensor=tracking_tensor,
-            img_cond_tensor=repaint_img_tensor,
-            prompt=prompt,
-            checkpoint_path=DEFAULT_MODEL_PATH
-        )
-        
-        return tracking_path, output_path
+        # 返回处理结果，但不应用跟踪
+        return tracking_path, video_tensor, tracking_tensor, repaint_img_tensor, fps
     except Exception as e:
         import traceback
         print(f"Processing failed: {str(e)}\n{traceback.format_exc()}")
-        return None, None
+        return None, None, None, None, None
 
 def process_camera_control(source, prompt, camera_motion, tracking_method):
     """Process camera control task"""
     try:
-        # Save uploaded files
+        # 保存上传的文件
         input_media_path = save_uploaded_file(source)
         if input_media_path is None:
-            return None, None
+            return None, None, None
         
         print(f"DEBUG: Camera motion: '{camera_motion}'")
         print(f"DEBUG: Tracking method: '{tracking_method}'")
@@ -317,24 +312,8 @@ def process_camera_control(source, prompt, camera_motion, tracking_method):
             # 使用在CPU上运行的cotracker
             pred_tracks, pred_visibility = generate_tracking_cotracker(video_tensor)
             
-            t, c, h, w = video_tensor.shape
-            new_width = 518
-            new_height = round(h * (new_width / w) / 14) * 14
-            resize_transform = transforms.Resize((new_height, new_width), interpolation=Image.BICUBIC)
-            video_vggt = resize_transform(video_tensor)  # [T, C, H, W]
-            
-            if new_height > 518:
-                start_y = (new_height - 518) // 2
-                video_vggt = video_vggt[:, :, start_y:start_y + 518, :]
-
-            vggt_model = get_vggt_model()
-
-            with torch.no_grad():
-                with torch.cuda.amp.autocast(dtype=das.dtype):
-                    video_vggt = video_vggt.unsqueeze(0)  # [1, T, C, H, W]
-                    aggregated_tokens_list, ps_idx = vggt_model.aggregator(video_vggt.to(das.device))
-                
-                    extr, intr = pose_encoding_to_extri_intri(vggt_model.camera_head(aggregated_tokens_list)[-1], video_vggt.shape[-2:])
+            # 使用封装的 VGGT 处理函数
+            extr, intr = process_vggt(video_tensor)
             
             cam_motion.set_intr(intr)
             cam_motion.set_extr(extr)
@@ -345,23 +324,15 @@ def process_camera_control(source, prompt, camera_motion, tracking_method):
                 pred_tracks = cam_motion.w2s_vggt(pred_tracks_world, extr, intr, poses) # [T, N, 3]
                 print("Camera motion applied")
             
-            tracking_path, tracking_tensor = das.visualize_tracking_cotracker(pred_tracks, None)
+            tracking_path, tracking_tensor = das.visualize_tracking_cotracker(pred_tracks, pred_visibility)
             print('Export tracking video via cotracker')
         
-        output_path = das.apply_tracking(
-            video_tensor=video_tensor,
-            fps=fps,  # 使用 load_media 返回的 fps
-            tracking_tensor=tracking_tensor,
-            img_cond_tensor=repaint_img_tensor,
-            prompt=prompt,
-            checkpoint_path=DEFAULT_MODEL_PATH
-        )
-        
-        return tracking_path, output_path
+        # 返回处理结果，但不应用跟踪
+        return tracking_path, video_tensor, tracking_tensor, repaint_img_tensor, fps
     except Exception as e:
         import traceback
         print(f"Processing failed: {str(e)}\n{traceback.format_exc()}")
-        return None, None
+        return None, None, None, None, None
 
 def process_object_manipulation(source, prompt, object_motion, object_mask, tracking_method):
     """Process object manipulation task"""
@@ -369,12 +340,12 @@ def process_object_manipulation(source, prompt, object_motion, object_mask, trac
         # Save uploaded files
         input_image_path = save_uploaded_file(source)
         if input_image_path is None:
-            return None, None
+            return None, None, None, None, None
             
         object_mask_path = save_uploaded_file(object_mask)
         if object_mask_path is None:
             print("Object mask not provided")
-            return None, None
+            return None, None, None, None, None
         
         das = get_das_pipeline()
         video_tensor, fps, is_video = load_media(input_image_path)
@@ -424,24 +395,8 @@ def process_object_manipulation(source, prompt, object_motion, object_mask, trac
             # 使用在CPU上运行的cotracker
             pred_tracks, pred_visibility = generate_tracking_cotracker(video_tensor)
             
-            t, c, h, w = video_tensor.shape
-            new_width = 518
-            new_height = round(h * (new_width / w) / 14) * 14
-            resize_transform = transforms.Resize((new_height, new_width), interpolation=Image.BICUBIC)
-            video_vggt = resize_transform(video_tensor)  # [T, C, H, W]
-            
-            if new_height > 518:
-                start_y = (new_height - 518) // 2
-                video_vggt = video_vggt[:, :, start_y:start_y + 518, :]
-
-            vggt_model = get_vggt_model()
-
-            with torch.no_grad():
-                with torch.cuda.amp.autocast(dtype=das.dtype):
-                    video_vggt = video_vggt.unsqueeze(0)  # [1, T, C, H, W]
-                    aggregated_tokens_list, ps_idx = vggt_model.aggregator(video_vggt.to(das.device))
-                
-                    extr, intr = pose_encoding_to_extri_intri(vggt_model.camera_head(aggregated_tokens_list)[-1], video_vggt.shape[-2:])
+            # 使用封装的 VGGT 处理函数
+            extr, intr = process_vggt(video_tensor)
             
             pred_tracks = motion_generator.apply_motion(
                 pred_tracks=pred_tracks.squeeze(),
@@ -453,23 +408,15 @@ def process_object_manipulation(source, prompt, object_motion, object_mask, trac
             )
             print(f"Object motion '{object_motion}' applied using provided mask")
             
-            tracking_path, tracking_tensor = das.visualize_tracking_cotracker(pred_tracks.unsqueeze(0), None)
+            tracking_path, tracking_tensor = das.visualize_tracking_cotracker(pred_tracks.unsqueeze(0), pred_visibility)
             print('Export tracking video via cotracker')
         
-        output_path = das.apply_tracking(
-            video_tensor=video_tensor,
-            fps=fps,  # 使用 load_media 返回的 fps
-            tracking_tensor=tracking_tensor,
-            img_cond_tensor=repaint_img_tensor,
-            prompt=prompt,
-            checkpoint_path=DEFAULT_MODEL_PATH
-        )
-        
-        return tracking_path, output_path
+        # 返回处理结果，但不应用跟踪
+        return tracking_path, video_tensor, tracking_tensor, repaint_img_tensor, fps
     except Exception as e:
         import traceback
         print(f"Processing failed: {str(e)}\n{traceback.format_exc()}")
-        return None, None
+        return None, None, None, None, None
 
 def process_mesh_animation(source, prompt, tracking_video, ma_repaint_option, ma_repaint_image):
     """Process mesh animation task"""
@@ -477,11 +424,11 @@ def process_mesh_animation(source, prompt, tracking_video, ma_repaint_option, ma
         # Save uploaded files
         input_video_path = save_uploaded_file(source)
         if input_video_path is None:
-            return None, None
+            return None, None, None, None, None
             
         tracking_video_path = save_uploaded_file(tracking_video)
         if tracking_video_path is None:
-            return None, None
+            return None, None, None, None, None
         
         das = get_das_pipeline()   
         video_tensor, fps, is_video = load_media(input_video_path)
@@ -494,7 +441,6 @@ def process_mesh_animation(source, prompt, tracking_video, ma_repaint_option, ma
             repaint_img_tensor, _, _ = load_media(repaint_path)
             repaint_img_tensor = repaint_img_tensor[0]  # 获取第一帧
         elif ma_repaint_option == "Yes":
-
             repainter = FirstFrameRepainter(gpu_id=GPU_ID, output_dir=OUTPUT_DIR)
             repaint_img_tensor = repainter.repaint(
                 video_tensor[0], 
@@ -502,20 +448,12 @@ def process_mesh_animation(source, prompt, tracking_video, ma_repaint_option, ma
                 depth_path=None
             )
 
-        output_path = das.apply_tracking(
-            video_tensor=video_tensor,
-            fps=fps,  # 使用 load_media 返回的 fps
-            tracking_tensor=tracking_tensor,
-            img_cond_tensor=repaint_img_tensor,
-            prompt=prompt,
-            checkpoint_path=DEFAULT_MODEL_PATH
-        )
-        
-        return tracking_video_path, output_path
+        # 直接返回上传的跟踪视频路径，而不是生成新的跟踪视频
+        return tracking_video_path, video_tensor, tracking_tensor, repaint_img_tensor, fps
     except Exception as e:
         import traceback
         print(f"Processing failed: {str(e)}\n{traceback.format_exc()}")
-        return None, None
+        return None, None, None, None, None
 
 def generate_tracking_cotracker(video_tensor, density=30):
     """在CPU上生成跟踪视频，只使用第一帧的深度信息，使用矩阵运算提高效率
@@ -569,22 +507,192 @@ def generate_tracking_cotracker(video_tensor, density=30):
     # 将结果返回
     return pred_tracks_with_depth.squeeze(0), pred_visibility.squeeze(0)
 
+@spaces.GPU(duration=240)
+def apply_tracking_unified(video_tensor, tracking_tensor, repaint_img_tensor, prompt, fps):
+    """统一的应用跟踪函数"""
+    try:
+        if video_tensor is None or tracking_tensor is None:
+            return None
+            
+        das = get_das_pipeline()
+        output_path = das.apply_tracking(
+            video_tensor=video_tensor,
+            fps=fps,
+            tracking_tensor=tracking_tensor,
+            img_cond_tensor=repaint_img_tensor,
+            prompt=prompt,
+            checkpoint_path=DEFAULT_MODEL_PATH
+        )
+        
+        print(f"生成的视频路径: {output_path}")
+        
+        # 确保返回的是绝对路径
+        if output_path and not os.path.isabs(output_path):
+            output_path = os.path.abspath(output_path)
+            
+        # 检查文件是否存在
+        if output_path and os.path.exists(output_path):
+            print(f"文件存在，大小: {os.path.getsize(output_path)} 字节")
+            return output_path
+        else:
+            print(f"警告: 输出文件不存在或路径无效: {output_path}")
+            return None
+    except Exception as e:
+        import traceback
+        print(f"Apply tracking failed: {str(e)}\n{traceback.format_exc()}")
+        return None
+
+# 添加在 apply_tracking_unified 函数之后，Gradio 界面定义之前
+
+def enable_apply_button(tracking_result):
+    """当跟踪视频生成后启用应用按钮"""
+    if tracking_result is not None:
+        return gr.update(interactive=True)
+    return gr.update(interactive=False)
+
+@spaces.GPU
+def process_vggt(video_tensor):
+    vggt_model = get_vggt_model()
+    
+    t, c, h, w = video_tensor.shape
+    new_width = 518
+    new_height = round(h * (new_width / w) / 14) * 14
+    resize_transform = transforms.Resize((new_height, new_width), interpolation=Image.BICUBIC)
+    video_vggt = resize_transform(video_tensor)  # [T, C, H, W]
+    
+    if new_height > 518:
+        start_y = (new_height - 518) // 2
+        video_vggt = video_vggt[:, :, start_y:start_y + 518, :]
+
+    with torch.no_grad():
+        with torch.cuda.amp.autocast(dtype=torch.float16):
+            video_vggt = video_vggt.unsqueeze(0)  # [1, T, C, H, W]
+            aggregated_tokens_list, ps_idx = vggt_model.aggregator(video_vggt.to("cuda"))
+        
+            extr, intr = pose_encoding_to_extri_intri(vggt_model.camera_head(aggregated_tokens_list)[-1], video_vggt.shape[-2:])
+    
+    return extr, intr 
+
+def load_examples():
+    """加载示例文件路径"""
+    samples_dir = os.path.join(project_root, "samples")
+    if not os.path.exists(samples_dir):
+        print(f"Warning: Samples directory not found at {samples_dir}")
+        return []
+    
+    examples_list = []
+    
+    # 为每个示例集创建一个示例项
+    # 示例1
+    example1 = [None] * 5  # [source, repaint_image, prompt, tracking_video, result_video]
+    for filename in os.listdir(samples_dir):
+        if filename.startswith("sample1_"):
+            if filename.endswith("_raw.mp4"):
+                example1[0] = os.path.join(samples_dir, filename)
+            elif filename.endswith("_repaint.png"):
+                example1[1] = os.path.join(samples_dir, filename)
+            elif filename.endswith("_tracking.mp4"):
+                example1[3] = os.path.join(samples_dir, filename)
+            elif filename.endswith("_result.mp4"):
+                example1[4] = os.path.join(samples_dir, filename)
+    
+    # 设置示例1的提示文本
+    example1[2] = "a rocket lifts off from the table and smoke erupt from its bottom."
+    
+    # 示例2
+    example2 = [None] * 5  # [source, repaint_image, prompt, tracking_video, result_video]
+    for filename in os.listdir(samples_dir):
+        if filename.startswith("sample2_"):
+            if filename.endswith("_raw.mp4"):
+                example2[0] = os.path.join(samples_dir, filename)
+            elif filename.endswith("_repaint.png"):
+                example2[1] = os.path.join(samples_dir, filename)
+            elif filename.endswith("_tracking.mp4"):
+                example2[3] = os.path.join(samples_dir, filename)
+            elif filename.endswith("_result.mp4"):
+                example2[4] = os.path.join(samples_dir, filename)
+    
+    # 设置示例2的提示文本
+    example2[2] = "A wonderful bright old-fasion red car is riding from left to right sun light is shining on the car, its reflection glittering. In the background is a deserted city in the noon, the roads and buildings are covered with green vegetation."
+    
+    # 添加示例到列表
+    if example1[0] is not None and example1[3] is not None:
+        examples_list.append(example1)
+    
+    if example2[0] is not None and example2[3] is not None:
+        examples_list.append(example2)
+    
+    # 添加其他示例（如果有）
+    sample_prefixes = set()
+    for filename in os.listdir(samples_dir):
+        if filename.endswith(('.mp4', '.png')):
+            prefix = filename.split('_')[0]
+            if prefix not in ["sample1", "sample2"]:
+                sample_prefixes.add(prefix)
+    
+    for prefix in sorted(sample_prefixes):
+        example = [None] * 5  # [source, repaint_image, prompt, tracking_video, result_video]
+        for filename in os.listdir(samples_dir):
+            if filename.startswith(f"{prefix}_"):
+                if filename.endswith("_raw.mp4"):
+                    example[0] = os.path.join(samples_dir, filename)
+                elif filename.endswith("_repaint.png"):
+                    example[1] = os.path.join(samples_dir, filename)
+                elif filename.endswith("_tracking.mp4"):
+                    example[3] = os.path.join(samples_dir, filename)
+                elif filename.endswith("_result.mp4"):
+                    example[4] = os.path.join(samples_dir, filename)
+        
+        # 添加默认提示文本
+        example[2] = "A beautiful scene"
+        
+        # 只有当至少有源文件和跟踪视频时才添加示例
+        if example[0] is not None and example[3] is not None:
+            examples_list.append(example)
+    
+    return examples_list
+
 # Create Gradio interface with updated layout
 with gr.Blocks(title="Diffusion as Shader") as demo:
     gr.Markdown("# Diffusion as Shader Web UI")
     gr.Markdown("### [Project Page](https://igl-hkust.github.io/das/) | [GitHub](https://github.com/IGL-HKUST/DiffusionAsShader)")
     
+    # 创建隐藏状态变量来存储中间结果
+    video_tensor_state = gr.State(None)
+    tracking_tensor_state = gr.State(None)
+    repaint_img_tensor_state = gr.State(None)
+    fps_state = gr.State(None)
+    
     with gr.Row():
         left_column = gr.Column(scale=1)
         right_column = gr.Column(scale=1)
 
     with right_column:
-        output_video = gr.Video(label="Generated Video")
         tracking_video = gr.Video(label="Tracking Video")
+        
+        # 初始状态下按钮不可用
+        apply_tracking_btn = gr.Button("Generate Video", variant="primary", size="lg", interactive=False)
+        output_video = gr.Video(label="Generated Video")
 
     with left_column:
-        source = gr.File(label="Source", file_types=["image", "video"])
-        common_prompt = gr.Textbox(label="Prompt", lines=2)
+        source_upload = gr.UploadButton("1. Upload Source", file_types=["image", "video"])
+        source_preview = gr.Video(label="Source Preview")
+        gr.Markdown("Upload a video or image, We will extract the motion and space structure from it")
+        
+        # 上传文件后更新预览
+        def update_source_preview(file):
+            if file is None:
+                return None
+            path = save_uploaded_file(file)
+            return path
+        
+        source_upload.upload(
+            fn=update_source_preview,
+            inputs=[source_upload],
+            outputs=[source_preview]
+        )
+        
+        common_prompt = gr.Textbox(label="2. Prompt: Describe the scene and the motion you want to create", lines=2)
         gr.Markdown(f"**Using GPU: {GPU_ID}**")
         
         with gr.Tabs() as task_tabs:
@@ -600,228 +708,308 @@ with gr.Blocks(title="Diffusion as Shader") as demo:
                         value="No"
                     )
                 gr.Markdown("### Note: If you want to use your own image as repainted first frame, please upload the image in below.")
-                # Custom image uploader (always visible)
-                mt_repaint_image = gr.File(
-                    label="Custom Repaint Image", 
-                    file_types=["image"]
+                
+                mt_repaint_upload = gr.UploadButton("3. Upload Repaint Image (Optional)", file_types=["image"])
+                mt_repaint_preview = gr.Image(label="Repaint Image Preview")
+                
+                # 上传文件后更新预览
+                mt_repaint_upload.upload(
+                    fn=update_source_preview,  # 复用相同的函数
+                    inputs=[mt_repaint_upload],
+                    outputs=[mt_repaint_preview]
                 )
                 
                 # Add run button for Motion Transfer tab
-                mt_run_btn = gr.Button("Run Motion Transfer", variant="primary", size="lg")
+                mt_run_btn = gr.Button("Generate Tracking", variant="primary", size="lg")
                 
-                # Connect to process function
+                # Connect to process function, but don't apply tracking
                 mt_run_btn.click(
                     fn=process_motion_transfer,
                     inputs=[
-                        source, common_prompt,
-                        mt_repaint_option, mt_repaint_image
+                        source_upload, common_prompt,
+                        mt_repaint_option, mt_repaint_upload
                     ],
-                    outputs=[tracking_video, output_video]
+                    outputs=[tracking_video, video_tensor_state, tracking_tensor_state, repaint_img_tensor_state, fps_state]
+                ).then(
+                    fn=enable_apply_button,
+                    inputs=[tracking_video],
+                    outputs=[apply_tracking_btn]
                 )
             
-            # Camera Control tab
-            with gr.TabItem("Camera Control"):
-                gr.Markdown("## Camera Control")
+            # # Camera Control tab
+            # with gr.TabItem("Camera Control"):
+            #     gr.Markdown("## Camera Control")
                 
-                cc_camera_motion = gr.Textbox(
-                    label="Current Camera Motion Sequence",
-                    placeholder="Your camera motion sequence will appear here...",
-                    interactive=False
-                )
+            #     cc_camera_motion = gr.Textbox(
+            #         label="Current Camera Motion Sequence",
+            #         placeholder="Your camera motion sequence will appear here...",
+            #         interactive=False
+            #     )
                 
-                # Use tabs for different motion types
-                with gr.Tabs() as cc_motion_tabs:
-                    # Translation tab
-                    with gr.TabItem("Translation (trans)"):
-                        with gr.Row():
-                            cc_trans_x = gr.Slider(minimum=-1.0, maximum=1.0, value=0.0, step=0.05, label="X-axis Movement")
-                            cc_trans_y = gr.Slider(minimum=-1.0, maximum=1.0, value=0.0, step=0.05, label="Y-axis Movement")
-                            cc_trans_z = gr.Slider(minimum=-1.0, maximum=1.0, value=0.0, step=0.05, label="Z-axis Movement (depth)")
+            #     # Use tabs for different motion types
+            #     with gr.Tabs() as cc_motion_tabs:
+            #         # Translation tab
+            #         with gr.TabItem("Translation (trans)"):
+            #             with gr.Row():
+            #                 cc_trans_x = gr.Slider(minimum=-1.0, maximum=1.0, value=0.0, step=0.05, label="X-axis Movement")
+            #                 cc_trans_y = gr.Slider(minimum=-1.0, maximum=1.0, value=0.0, step=0.05, label="Y-axis Movement")
+            #                 cc_trans_z = gr.Slider(minimum=-1.0, maximum=1.0, value=0.0, step=0.05, label="Z-axis Movement (depth)")
                         
-                        with gr.Row():
-                            cc_trans_start = gr.Number(minimum=0, maximum=48, value=0, step=1, label="Start Frame", precision=0)
-                            cc_trans_end = gr.Number(minimum=0, maximum=48, value=48, step=1, label="End Frame", precision=0)
+            #             with gr.Row():
+            #                 cc_trans_start = gr.Number(minimum=0, maximum=48, value=0, step=1, label="Start Frame", precision=0)
+            #                 cc_trans_end = gr.Number(minimum=0, maximum=48, value=48, step=1, label="End Frame", precision=0)
                         
-                        cc_trans_note = gr.Markdown("""
-                        **Translation Notes:**
-                        - Positive X: Move right, Negative X: Move left
-                        - Positive Y: Move down, Negative Y: Move up
-                        - Positive Z: Zoom in, Negative Z: Zoom out
-                        """)
+            #             cc_trans_note = gr.Markdown("""
+            #             **Translation Notes:**
+            #             - Positive X: Move right, Negative X: Move left
+            #             - Positive Y: Move down, Negative Y: Move up
+            #             - Positive Z: Zoom in, Negative Z: Zoom out
+            #             """)
                         
-                        # Add translation button in the Translation tab
-                        cc_add_trans = gr.Button("Add Camera Translation", variant="secondary")
+            #             # Add translation button in the Translation tab
+            #             cc_add_trans = gr.Button("Add Camera Translation", variant="secondary")
                         
-                        # Function to add translation motion
-                        def add_translation_motion(current_motion, trans_x, trans_y, trans_z, trans_start, trans_end):
-                            # Format: trans dx dy dz [start_frame end_frame]
-                            frame_range = f" {int(trans_start)} {int(trans_end)}" if trans_start != 0 or trans_end != 48 else ""
-                            new_motion = f"trans {trans_x:.2f} {trans_y:.2f} {trans_z:.2f}{frame_range}"
+            #             # Function to add translation motion
+            #             def add_translation_motion(current_motion, trans_x, trans_y, trans_z, trans_start, trans_end):
+            #                 # Format: trans dx dy dz [start_frame end_frame]
+            #                 frame_range = f" {int(trans_start)} {int(trans_end)}" if trans_start != 0 or trans_end != 48 else ""
+            #                 new_motion = f"trans {trans_x:.2f} {trans_y:.2f} {trans_z:.2f}{frame_range}"
                             
-                            # Append to existing motion string with semicolon separator if needed
-                            if current_motion and current_motion.strip():
-                                updated_motion = f"{current_motion}; {new_motion}"
-                            else:
-                                updated_motion = new_motion
+            #                 # Append to existing motion string with semicolon separator if needed
+            #                 if current_motion and current_motion.strip():
+            #                     updated_motion = f"{current_motion}; {new_motion}"
+            #                 else:
+            #                     updated_motion = new_motion
                             
-                            return updated_motion
+            #                 return updated_motion
                         
-                        # Connect translation button
-                        cc_add_trans.click(
-                            fn=add_translation_motion,
-                            inputs=[
-                                cc_camera_motion,
-                                cc_trans_x, cc_trans_y, cc_trans_z, cc_trans_start, cc_trans_end
-                            ],
-                            outputs=[cc_camera_motion]
-                        )
+            #             # Connect translation button
+            #             cc_add_trans.click(
+            #                 fn=add_translation_motion,
+            #                 inputs=[
+            #                     cc_camera_motion,
+            #                     cc_trans_x, cc_trans_y, cc_trans_z, cc_trans_start, cc_trans_end
+            #                 ],
+            #                 outputs=[cc_camera_motion]
+            #             )
                     
-                    # Rotation tab
-                    with gr.TabItem("Rotation (rot)"):
-                        with gr.Row():
-                            cc_rot_axis = gr.Dropdown(choices=["x", "y", "z"], value="y", label="Rotation Axis")
-                            cc_rot_angle = gr.Slider(minimum=-30, maximum=30, value=5, step=1, label="Rotation Angle (degrees)")
+            #         # Rotation tab
+            #         with gr.TabItem("Rotation (rot)"):
+            #             with gr.Row():
+            #                 cc_rot_axis = gr.Dropdown(choices=["x", "y", "z"], value="y", label="Rotation Axis")
+            #                 cc_rot_angle = gr.Slider(minimum=-30, maximum=30, value=5, step=1, label="Rotation Angle (degrees)")
                         
-                        with gr.Row():
-                            cc_rot_start = gr.Number(minimum=0, maximum=48, value=0, step=1, label="Start Frame", precision=0)
-                            cc_rot_end = gr.Number(minimum=0, maximum=48, value=48, step=1, label="End Frame", precision=0)
+            #             with gr.Row():
+            #                 cc_rot_start = gr.Number(minimum=0, maximum=48, value=0, step=1, label="Start Frame", precision=0)
+            #                 cc_rot_end = gr.Number(minimum=0, maximum=48, value=48, step=1, label="End Frame", precision=0)
                         
-                        cc_rot_note = gr.Markdown("""
-                        **Rotation Notes:**
-                        - X-axis rotation: Tilt camera up/down
-                        - Y-axis rotation: Pan camera left/right
-                        - Z-axis rotation: Roll camera
-                        """)
+            #             cc_rot_note = gr.Markdown("""
+            #             **Rotation Notes:**
+            #             - X-axis rotation: Tilt camera up/down
+            #             - Y-axis rotation: Pan camera left/right
+            #             - Z-axis rotation: Roll camera
+            #             """)
                         
-                        # Add rotation button in the Rotation tab
-                        cc_add_rot = gr.Button("Add Camera Rotation", variant="secondary")
+            #             # Add rotation button in the Rotation tab
+            #             cc_add_rot = gr.Button("Add Camera Rotation", variant="secondary")
                         
-                        # Function to add rotation motion
-                        def add_rotation_motion(current_motion, rot_axis, rot_angle, rot_start, rot_end):
-                            # Format: rot axis angle [start_frame end_frame]
-                            frame_range = f" {int(rot_start)} {int(rot_end)}" if rot_start != 0 or rot_end != 48 else ""
-                            new_motion = f"rot {rot_axis} {rot_angle}{frame_range}"
+            #             # Function to add rotation motion
+            #             def add_rotation_motion(current_motion, rot_axis, rot_angle, rot_start, rot_end):
+            #                 # Format: rot axis angle [start_frame end_frame]
+            #                 frame_range = f" {int(rot_start)} {int(rot_end)}" if rot_start != 0 or rot_end != 48 else ""
+            #                 new_motion = f"rot {rot_axis} {rot_angle}{frame_range}"
                             
-                            # Append to existing motion string with semicolon separator if needed
-                            if current_motion and current_motion.strip():
-                                updated_motion = f"{current_motion}; {new_motion}"
-                            else:
-                                updated_motion = new_motion
+            #                 # Append to existing motion string with semicolon separator if needed
+            #                 if current_motion and current_motion.strip():
+            #                     updated_motion = f"{current_motion}; {new_motion}"
+            #                 else:
+            #                     updated_motion = new_motion
                             
-                            return updated_motion
+            #                 return updated_motion
                         
-                        # Connect rotation button
-                        cc_add_rot.click(
-                            fn=add_rotation_motion,
-                            inputs=[
-                                cc_camera_motion,
-                                cc_rot_axis, cc_rot_angle, cc_rot_start, cc_rot_end
-                            ],
-                            outputs=[cc_camera_motion]
-                        )
+            #             # Connect rotation button
+            #             cc_add_rot.click(
+            #                 fn=add_rotation_motion,
+            #                 inputs=[
+            #                     cc_camera_motion,
+            #                     cc_rot_axis, cc_rot_angle, cc_rot_start, cc_rot_end
+            #                 ],
+            #                 outputs=[cc_camera_motion]
+            #             )
                 
-                # Add a clear button to reset the motion sequence
-                cc_clear_motion = gr.Button("Clear All Motions", variant="stop")
+            #     # Add a clear button to reset the motion sequence
+            #     cc_clear_motion = gr.Button("Clear All Motions", variant="stop")
                 
-                def clear_camera_motion():
-                    return ""
+            #     def clear_camera_motion():
+            #         return ""
                 
-                cc_clear_motion.click(
-                    fn=clear_camera_motion,
-                    inputs=[],
-                    outputs=[cc_camera_motion]
-                )
-
-                cc_tracking_method = gr.Radio(
-                    label="Tracking Method",
-                    choices=["moge", "cotracker"],
-                    value="cotracker"
-                )
+            #     cc_clear_motion.click(
+            #         fn=clear_camera_motion,
+            #         inputs=[],
+            #         outputs=[cc_camera_motion]
+            #     )
+
+            #     cc_tracking_method = gr.Radio(
+            #         label="Tracking Method",
+            #         choices=["moge", "cotracker"],
+            #         value="cotracker"
+            #     )
                 
-                # Add run button for Camera Control tab
-                cc_run_btn = gr.Button("Run Camera Control", variant="primary", size="lg")
+            #     # Add run button for Camera Control tab
+            #     cc_run_btn = gr.Button("Generate Tracking", variant="primary", size="lg")
                 
-                # Connect to process function
-                cc_run_btn.click(
-                    fn=process_camera_control,
-                    inputs=[
-                        source, common_prompt,
-                        cc_camera_motion, cc_tracking_method
-                    ],
-                    outputs=[tracking_video, output_video]
-                )
+            #     # Connect to process function, but don't apply tracking
+            #     cc_run_btn.click(
+            #         fn=process_camera_control,
+            #         inputs=[
+            #             source_upload, common_prompt,
+            #             cc_camera_motion, cc_tracking_method
+            #         ],
+            #         outputs=[tracking_video, video_tensor_state, tracking_tensor_state, repaint_img_tensor_state, fps_state]
+            #     ).then(
+            #         fn=enable_apply_button,
+            #         inputs=[tracking_video],
+            #         outputs=[apply_tracking_btn]
+            #     )
             
-            # Object Manipulation tab
-            with gr.TabItem("Object Manipulation"):
-                gr.Markdown("## Object Manipulation")
-                om_object_mask = gr.File(
-                    label="Object Mask Image", 
-                    file_types=["image"]
-                )
-                gr.Markdown("Upload a binary mask image, white areas indicate the object to manipulate")
-                om_object_motion = gr.Dropdown(
-                    label="Object Motion Type",
-                    choices=["up", "down", "left", "right", "front", "back", "rot"],
-                    value="up"
-                )
-                om_tracking_method = gr.Radio(
-                    label="Tracking Method",
-                    choices=["moge", "cotracker"],
-                    value="cotracker"
-                )
+            # # Object Manipulation tab
+            # with gr.TabItem("Object Manipulation"):
+            #     gr.Markdown("## Object Manipulation")
+            #     om_object_mask = gr.File(
+            #         label="Object Mask Image", 
+            #         file_types=["image"]
+            #     )
+            #     gr.Markdown("Upload a binary mask image, white areas indicate the object to manipulate")
+            #     om_object_motion = gr.Dropdown(
+            #         label="Object Motion Type",
+            #         choices=["up", "down", "left", "right", "front", "back", "rot"],
+            #         value="up"
+            #     )
+            #     om_tracking_method = gr.Radio(
+            #         label="Tracking Method",
+            #         choices=["moge", "cotracker"],
+            #         value="cotracker"
+            #     )
                 
-                # Add run button for Object Manipulation tab
-                om_run_btn = gr.Button("Run Object Manipulation", variant="primary", size="lg")
+            #     # Add run button for Object Manipulation tab
+            #     om_run_btn = gr.Button("Generate Tracking", variant="primary", size="lg")
                 
-                # Connect to process function
-                om_run_btn.click(
-                    fn=process_object_manipulation,
-                    inputs=[
-                        source, common_prompt,
-                        om_object_motion, om_object_mask, om_tracking_method
-                    ],
-                    outputs=[tracking_video, output_video]
-                )
+            #     # Connect to process function, but don't apply tracking
+            #     om_run_btn.click(
+            #         fn=process_object_manipulation,
+            #         inputs=[
+            #             source_upload, common_prompt,
+            #             om_object_motion, om_object_mask, om_tracking_method
+            #         ],
+            #         outputs=[tracking_video, video_tensor_state, tracking_tensor_state, repaint_img_tensor_state, fps_state]
+            #     ).then(
+            #         fn=enable_apply_button,
+            #         inputs=[tracking_video],
+            #         outputs=[apply_tracking_btn]
+            #     )
             
-            # Animating meshes to video tab
-            with gr.TabItem("Animating meshes to video"):
-                gr.Markdown("## Mesh Animation to Video")
-                gr.Markdown("""
-                    Note: Currently only supports tracking videos generated with Blender (version > 4.0).
-                    Please run the script `scripts/blender.py` in your Blender project to generate tracking videos.
-                """)
-                ma_tracking_video = gr.File(
-                    label="Tracking Video",
-                    file_types=["video"]
-                )
-                gr.Markdown("Tracking video needs to be generated from Blender")
+            # # Animating meshes to video tab
+            # with gr.TabItem("Animating meshes to video"):
+            #     gr.Markdown("## Mesh Animation to Video")
+            #     gr.Markdown("""
+            #         Note: Currently only supports tracking videos generated with Blender (version > 4.0).
+            #         Please run the script `scripts/blender.py` in your Blender project to generate tracking videos.
+            #     """)
+            #     ma_tracking_video = gr.File(
+            #         label="Tracking Video",
+            #         file_types=["video"],
+            #         # 添加 change 事件处理器，当上传文件时自动激活 Generate Video 按钮
+            #         elem_id="ma_tracking_video"
+            #     )
+            #     gr.Markdown("Tracking video needs to be generated from Blender")
                 
-                # Simplified controls - Radio buttons for Yes/No and separate file upload
-                with gr.Row():
-                    ma_repaint_option = gr.Radio(
-                        label="Repaint First Frame",
-                        choices=["No", "Yes"],
-                        value="No"
-                    )
-                gr.Markdown("### Note: If you want to use your own image as repainted first frame, please upload the image in below.")
-                # Custom image uploader (always visible)
-                ma_repaint_image = gr.File(
-                    label="Custom Repaint Image", 
-                    file_types=["image"]
-                )
+            #     # Simplified controls - Radio buttons for Yes/No and separate file upload
+            #     with gr.Row():
+            #         ma_repaint_option = gr.Radio(
+            #             label="Repaint First Frame",
+            #             choices=["No", "Yes"],
+            #             value="No"
+            #         )
+            #     gr.Markdown("### Note: If you want to use your own image as repainted first frame, please upload the image in below.")
+            #     # Custom image uploader (always visible)
+            #     ma_repaint_image = gr.File(
+            #         label="Custom Repaint Image", 
+            #         file_types=["image"]
+            #     )
                 
-                # Add run button for Mesh Animation tab
-                ma_run_btn = gr.Button("Run Mesh Animation", variant="primary", size="lg")
+            #     # 修改按钮名称为 "Apply Repaint"
+            #     ma_run_btn = gr.Button("Apply Repaint", variant="primary", size="lg")
                 
-                # Connect to process function
-                ma_run_btn.click(
-                    fn=process_mesh_animation,
-                    inputs=[
-                        source, common_prompt,
-                        ma_tracking_video, ma_repaint_option, ma_repaint_image
-                    ],
-                    outputs=[tracking_video, output_video]
-                )
+            #     # 添加 tracking video 上传事件处理
+            #     def handle_tracking_upload(file):
+            #         if file is not None:
+            #             tracking_path = save_uploaded_file(file)
+            #             if tracking_path:
+            #                 return tracking_path, gr.update(interactive=True)
+            #         return None, gr.update(interactive=False)
+                
+            #     # 当上传 tracking video 时，直接显示并激活 Generate Video 按钮
+            #     ma_tracking_video.change(
+            #         fn=handle_tracking_upload,
+            #         inputs=[ma_tracking_video],
+            #         outputs=[tracking_video, apply_tracking_btn]
+            #     )
+                
+            #     # 修改 process_mesh_animation 函数的行为
+            #     def process_mesh_animation_repaint(source, prompt, ma_repaint_option, ma_repaint_image):
+            #         """只处理重绘部分，不处理跟踪视频"""
+            #         try:
+            #             # 保存上传的文件
+            #             input_video_path = save_uploaded_file(source)
+            #             if input_video_path is None:
+            #                 return None, None, None, None
+                        
+            #             das = get_das_pipeline()   
+            #             video_tensor, fps, is_video = load_media(input_video_path)
+            #             das.fps = fps
+                        
+            #             repaint_img_tensor = None
+            #             if ma_repaint_image is not None:
+            #                 repaint_path = save_uploaded_file(ma_repaint_image)
+            #                 repaint_img_tensor, _, _ = load_media(repaint_path)
+            #                 repaint_img_tensor = repaint_img_tensor[0]
+            #             elif ma_repaint_option == "Yes":
+            #                 repainter = FirstFrameRepainter(gpu_id=GPU_ID, output_dir=OUTPUT_DIR)
+            #                 repaint_img_tensor = repainter.repaint(
+            #                     video_tensor[0], 
+            #                     prompt=prompt,
+            #                     depth_path=None
+            #                 )
+                        
+            #             # 返回处理结果，但不包括跟踪视频路径
+            #             return video_tensor, None, repaint_img_tensor, fps
+            #         except Exception as e:
+            #             import traceback
+            #             print(f"Processing failed: {str(e)}\n{traceback.format_exc()}")
+            #             return None, None, None, None
+                
+            #     # 连接到修改后的处理函数
+            #     ma_run_btn.click(
+            #         fn=process_mesh_animation_repaint,
+            #         inputs=[
+            #             source_upload, common_prompt,
+            #             ma_repaint_option, ma_repaint_image
+            #         ],
+            #         outputs=[video_tensor_state, tracking_tensor_state, repaint_img_tensor_state, fps_state]
+            #     )
+        
+    # 在所有 UI 元素定义之后，添加 Examples 组件
+    examples_list = load_examples()
+    if examples_list:
+        with gr.Blocks() as examples_block:
+            gr.Examples(
+                examples=examples_list,
+                inputs=[source_preview, mt_repaint_preview, common_prompt, tracking_video, output_video],
+                outputs=[source_preview, mt_repaint_preview, common_prompt, tracking_video, output_video],
+                fn=lambda *args: args,  # 简单地返回输入作为输出
+                cache_examples=True,
+                label="Examples"
+            )
 
 # Launch interface
 if __name__ == "__main__":
@@ -831,4 +1019,4 @@ if __name__ == "__main__":
         print("Creating public link for remote access")
     
     # Launch interface
-    demo.launch(share=args.share, server_port=args.port) 
+    demo.launch(share=args.share, server_port=args.port)