ProPainter
ProPainter copied to clipboard
sczhou
The result is not very good, do you want to train it?
I also have the same issue when trying to remove people from videos, the removed areas come out as very pixelated or very blurred
This is the result I got from the walking video from the Davis dataset. I feel the result would be a lot better if somehow the inpainted areas aren't so blurry.
https://github.com/sczhou/ProPainter/assets/54892636/f3dd3898-2df8-4048-a937-6d4a936fbac6
https://github.com/sczhou/ProPainter/assets/54892636/9190f0ea-e795-4514-bf1c-0981dbbeec21
Results are better when you have clean background areas in other frames - either the people/objects have to move, or the camera has to move to reveal them . If the mask area obscures everything, and there are no clean textures behind the "object/people" in nearby frames - you cannot benefit from using other frames to inpaint
https://github.com/sczhou/ProPainter/assets/1106791/4260b1ea-573d-42bb-849f-10f3fac2ddbe
https://github.com/sczhou/ProPainter/assets/1106791/1cd48a7d-24cf-42f4-9e2d-6cbce39a1c55
Results are better when you have clean background areas in other frames - either the people/objects have to move, or the camera has to move to reveal them . If the mask area obscures everything, and there are no clean textures behind the "object/people" in nearby frames - you cannot benefit from using other frames to inpaint
Results are better when you have clean background areas in other frames - either the people/objects have to move, or the camera has to move to reveal them . If the mask area obscures everything, and there are no clean textures behind the "object/people" in nearby frames - you cannot benefit from using other frames to inpaint
Indeed, the current results in removing watermarks or text are decent, using only a pre-trained model.
But if you look at the FGT demo video or try to run the exact same video using FGT, the results are a lot cleaner with the same Davis walking example for some reason. Is it because Pro Painter is based on SSTN that the output results are the same?
https://www.youtube.com/watch?v=BC32n-NncPs
can you tell me what tool you used to make mask?Do I need to draw the mask one frame by one frame?
can you tell me what tool you used to make mask?Do I need to draw the mask one frame by one frame?
sam-pt the project
Hello, can you help me, I encountered a bug: FileNotFoundError: [Errno 2] No such file or directory: '/mnt/lustre/sczhou/VQGANs/CodeMOVI/experiments_model/recurrent_flow_completion_v5_train_flowcomp_v5/gen_760000.pth' how to get the weight file of recurrent flow completion.
