The results were relatively bad? Why?

[lzwRover321]

Thanks for the excellent code, when I inference the picture. gt: mask: Lama result:(The result of Lama https://github.com/advimman/lama) PUT result:(The result of model tpami2024_vit_base_naturalscene_res512) python scripts/inference.py --func inference_inpainting \ --name OUTPUT/tpami2024_vit_base_naturalscene_res512/checkpoint/000599e_742800iter.pth \ --input_res 512,512 \ --num_token_per_iter -1 \ --num_token_for_sampling 1 \ --num_sample 1 \ --num_replicate 1 \ --image_dir data/naturalscene_512_sample/gt \ --mask_dir data/naturalscene_512_sample/mr0.5_0.6 \ --gpu 0

[liuqk3]

For the model trained on Places, You can try the following two commands:

python scripts/inference.py --func inference_inpainting
--name OUTPUT/tpami2024_vit_base_naturalscene_res512/checkpoint/000599e_742800iter.pth
--input_res 512,512
--num_token_per_iter 20
--num_token_for_sampling 200
--num_sample 1
--num_replicate 1
--image_dir data/naturalscene_512_sample/gt
--mask_dir data/naturalscene_512_sample/mr0.5_0.6
--gpu 0

# or

python scripts/inference.py --func inference_inpainting
--name OUTPUT/tpami2024_vit_base_naturalscene_res512/checkpoint/000599e_742800iter.pth
--input_res 512,512
--num_token_per_iter -1
--num_token_for_sampling 20
--num_sample 1
--num_replicate 1
--image_dir data/naturalscene_512_sample/gt
--mask_dir data/naturalscene_512_sample/mr0.5_0.6
--gpu 0

[lzwRover321]

naturalscene_512_sample

Thank you for your reply, but I still haven't got a good result.

[liuqk3]

The mask should be in PNG format. The pixel values in mask should be 0 or 255. The JPEG mask image will introduce artifacts.

[lzwRover321]

The mask should be in PNG format. The pixel values in mask should be 0 or 255. The JPEG mask image will introduce artifacts.

Okay, so the PNG mask is a single channel. I got better results, but the edges don't seem to transition smoothly. Great work, thanks again!