Diffusion process may induce low-quality images, and harmful artifacts or regions. In this stage, we use clip directional similarity metric and segmentation models to gradually filter out harmful synthetic samples and artifact regions.
Please run the following command to filter out low-quality images.
cd filter
python filter.py --image-dir-1 <> --image-dir-2 <> --save-image-dir <> --dis-image-dir <>> --text-path-1 <>> --text-path-2 <>
--image-dir-1
and --image-dir-2
are directories to original and synthetic images.
Then the high-quality images will be stored in --save-image-dir
, low-quality images will be filtered out in --dis-image-dir
Please run the following command to detect artifact regions with segmentation model.
python detector.py --cnofig <config> --checkpoint <checkpoint> --meta_file_path <> --real-img-path <> --real-label-path <> --syn-img-path <> --object-mask-path <> --filtered-label-path <>
We use the classical pre-trained Upernet-R101 model to calculate per-pixel loss on real images and synthetic image, and then filter out noisy synthetic regions. At last, the filtered semseg label will be stored in --filtered-label-path