Update web-demos/hugging_face/app.py

web-demos/hugging_face/app.py

@@ -297,40 +297,64 @@ def vos_tracking_video(video_state, interactive_state, mask_dropdown):
     return video_output, video_state, interactive_state, operation_log, operation_log
 
 # inpaint 
-def inpaint_video(video_state, resize_ratio_number, dilate_radius_number, raft_iter_number, subvideo_length_number, neighbor_length_number, ref_stride_number, mask_dropdown):
-    operation_log = [("",""), ("Inpainting finished!","Normal")]
+def inpaint_video(video_state, *_args):
+    operation_log = [("",""), ("Inpainting started in safe high-quality mode.","Normal")]
 
+    # Получаем кадры и маски
     frames = np.asarray(video_state["origin_images"])
     fps = video_state["fps"]
     inpaint_masks = np.asarray(video_state["masks"])
+
+    # Маскировка (если не выбраны вручную)
+    mask_dropdown = _args[-1]
     if len(mask_dropdown) == 0:
         mask_dropdown = ["mask_001"]
     mask_dropdown.sort()
-    # convert mask_dropdown to mask numbers
-    inpaint_mask_numbers = [int(mask_dropdown[i].split("_")[1]) for i in range(len(mask_dropdown))]
-    # interate through all masks and remove the masks that are not in mask_dropdown
-    unique_masks = np.unique(inpaint_masks)
-    num_masks = len(unique_masks) - 1
-    for i in range(1, num_masks + 1):
-        if i in inpaint_mask_numbers:
-            continue
-        inpaint_masks[inpaint_masks==i] = 0
-    
-    # inpaint for videos
-    inpainted_frames = model.baseinpainter.inpaint(frames, 
-                                                   inpaint_masks, 
-                                                   ratio=resize_ratio_number, 
-                                                   dilate_radius=dilate_radius_number,
-                                                   raft_iter=raft_iter_number,
-                                                   subvideo_length=subvideo_length_number, 
-                                                   neighbor_length=neighbor_length_number, 
-                                                   ref_stride=ref_stride_number)   # numpy array, T, H, W, 3
-
+    inpaint_mask_numbers = [int(name.split("_")[1]) for name in mask_dropdown]
+    for i in range(1, np.max(inpaint_masks) + 1):
+        if i not in inpaint_mask_numbers:
+            inpaint_masks[inpaint_masks == i] = 0
+
+    # Фиксированные безопасные настройки
+    chunk_size = 20              # обрабатываем по 20 кадров
+    fixed_resize_ratio = 1.0     # не уменьшаем разрешение
+    fixed_dilate_radius = 4
+    fixed_raft_iter = 20
+    fixed_neighbor_length = 5
+    fixed_ref_stride = 10
+
+    total_len = len(frames)
+    inpainted_all = []
+
+    for start in range(0, total_len, chunk_size):
+        end = min(start + chunk_size, total_len)
+        print(f"Inpainting chunk {start} → {end}")
+        chunk_frames = frames[start:end]
+        chunk_masks = inpaint_masks[start:end]
+
+        # Инференс inpainting
+        chunk_result = model.baseinpainter.inpaint(
+            chunk_frames,
+            chunk_masks,
+            ratio=fixed_resize_ratio,
+            dilate_radius=fixed_dilate_radius,
+            raft_iter=fixed_raft_iter,
+            subvideo_length=chunk_size,
+            neighbor_length=fixed_neighbor_length,
+            ref_stride=fixed_ref_stride
+        )
+        inpainted_all.extend(chunk_result)
+
+    # Сохраняем с повышенным битрейтом
     output_path = "./result/inpaint/{}".format(video_state["video_name"])
-    video_output = generate_video_from_frames(inpainted_frames, output_path=output_path, fps=fps, bitrate="30M") # import video_input to name the output video
+    video_output = generate_video_from_frames(
+        inpainted_all,
+        output_path=output_path,
+        fps=fps,
+        bitrate="30M"
+    )
 
     return video_output, operation_log, operation_log
-    print("Inpainting resolution:", inpainted_frames[0].shape)
 
 
 # generate video after vos inference