IOPaint
IOPaint copied to clipboard
Sanster
Do you support custom MASK?
https://github.com/Sanster/lama-cleaner/releases/tag/0.26.0
Test use assets/dog.jpg, by "background".
Hi, Is there anything I need to pay attention about mask images? I use the mask image from :
segment-anything
it remove dog success , but it filled with other object there.
i find the mask image from [segment-anything] is a grayscale image, so i convert it to RGB, but it still works wrong.
And if i use mak image from [lama-cleaner] download, it works very well.
https://github.com/Sanster/lama-cleaner/releases/tag/0.26.0
Test use assets/dog.jpg, by "background". Hi, Is there anything I need to pay attention about mask images? I use the mask image from : segment-anything
it works wrong, remove dog success , but it filled with other object there. i find the mask image from [segment-anything] is a grayscale image, so i convert it to RGB, but it still works wrong.
maybe you can show some results to compare
@SunSoul Can you share a picture of your results? If you directly use the results of segment-anything, the LaMa model's performance is indeed not good(The mask should be slightly larger than the target area to be inpainted.). To achieve better results, you need to perform a dilation operation on the mask result of segment-anything. https://github.com/Sanster/lama-cleaner/blob/37e9137ac70ef9021224a7c3037b21d35b835a01/lama_cleaner/plugins/interactive_seg.py#L68
By the way, segment-anything is already supported in lama-cleaner 1.1.1, here is a demo video. You can start server with following command line args:
--enable-interactive-seg --interactive-seg-model=vit_l --interactive-seg-device=cpu
@SunSoul Can you share a picture of your results? If you directly use the results of segment-anything, the LaMa model's performance is indeed not good(The mask should be slightly larger than the target area to be inpainted.). To achieve better results, you need to perform a dilation operation on the mask result of segment-anything.
https://github.com/Sanster/lama-cleaner/blob/37e9137ac70ef9021224a7c3037b21d35b835a01/lama_cleaner/plugins/interactive_seg.py#L68
By the way, segment-anything is already supported in lama-cleaner 1.1.1, here is a demo video. You can start server with following command line args:
--enable-interactive-seg --interactive-seg-model=vit_l --interactive-seg-device=cpu
Just now I was re-editing the issue! I'll test it, thanks!
https://github.com/Sanster/lama-cleaner/releases/tag/0.26.0
Test use assets/dog.jpg, by "background". Hi, Is there anything I need to pay attention about mask images? I use the mask image from : segment-anything it works wrong, remove dog success , but it filled with other object there. i find the mask image from [segment-anything] is a grayscale image, so i convert it to RGB, but it still works wrong.
maybe you can show some results to compare
Just now I was re-editing the issue, thanks!