PSNR and SSIM for Image Restoration

[sdj22-beep]

Hi, thanks for the amazing work! I am interested in reproducing the restoration results of PixWizard. I tried the method using inference_pixwizard.sh on samples from the SOTS dataset. However, the PSNR and SSIM I am getting are different from the paper and are quite low: 21.28dB, 0.738. I also tried the script for ./infer_examples/fig4.jpg. The PSNR and SSIM are very low for that image as well: 16.34dB, 0.567.

I think this is because of the SDXL VAE as there are structural distortions in the restored images. I have attached the results below for ./infer_examples/fig4.jpg and an image from SOTS below (second image is ground truth in both cases). Could you please help me resolve this issue?

Thanks!

[AFeng-x]

Hi there~ Thank you for your question. The differences in PSNR and SSIM may be influenced by the following factors:

1. Image resolution
2. Sampling steps
3. The currently released model checkpoint is a trial version for multi-task balance, which might affect the performance (either higher or lower) on certain tasks.