fix bug

dkt/pipelines/pipeline.py

@@ -958,8 +958,12 @@ class DKTPipeline:
             cfg_scale=1.0,
         )
         print(f"[P4] main_pipe done, video length={len(video)}")
+        
+        del vae_outs
+        video = video.cpu()
         torch.cuda.empty_cache()
 
+
         processed_video = video[:frame_length]
         processed_video = [resize_frame(frame, original_height, original_width) for frame in processed_video]
 
@@ -991,7 +995,7 @@ class DKTPipeline:
 
 
 
-        if vis_pc and self.prompt == 'depth':
+        if False and vis_pc and self.prompt == 'depth':
             vis_pc_num = 4
             indices = np.linspace(0, frame_length-1, vis_pc_num)
             indices = np.round(indices).astype(np.int32)
@@ -1001,7 +1005,8 @@ class DKTPipeline:
             return_dict['rgb_frames'] = origin_frames
 
 
-        if get_moge_intrinsics:
+        if  False and get_moge_intrinsics :
+            
             demo_idx = 0
             origin_frames_demo = origin_frames[demo_idx]
             prediced_depth_map_np_demo = prediced_depth_map_np[demo_idx]
@@ -1016,7 +1021,6 @@ class DKTPipeline:
             metric_depth, scale, shift = transfer_pred_disp2depth(prediced_depth_map_np_demo, moge_depth, moge_mask, return_scale_shift=True)
 
 
-
             moge_intrinsics[0, 0] *= original_width 
             moge_intrinsics[1, 1] *= original_height
             moge_intrinsics[0, 2] *= original_width

examples/16.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4ca7cbbd94cf56db8d7f70e9e1092c9b69cb6c6a9799cb3e021a1f722553e978
-size 561537
+oid sha256:38266a7ec176347f3890a58881b8b4d388b607c5af05c97d14220224a9dda7b2
+size 1140628

examples/17.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4919a4adbace8e8b4f6cdb481b73a342d099b1912392f71058d9cc36c6e1ad93
-size 689947
+oid sha256:5534789b249fb9050bc9eb4df11e6e14a7b6f14252937c5cce27104e16c9bf10
+size 1211725