Regarding the issues with the kid metric and the psnr metric.

yuhaoliu7456 / Diff-Plugin

[CVPR 2024] Official code release of our paper "Diff-Plugin: Revitalizing Details for Diffusion-based Low-level tasks"

130 stars 4 forks source link

Regarding the issues with the kid metric and the psnr metric. #17

Open flyingfishing opened 1 month ago

flyingfishing commented 1 month ago

Dear Author,

I have encountered an issue while replicating your highlight task using the shiq dataset. I noticed that the fid metric matches, but there is a significant discrepancy in the kid metric. Have you ever encountered this situation? Additionally, I used psnr and ssim metrics for evaluation and found that the psnr metric was only 18. Did you attempt to use the psnr metric during your selection of evaluation metrics?

yuhaoliu7456 commented 3 weeks ago

Did you re-train the model or load our weights and re-test them? Can you provide more information? I did not encounter this issue before. As for the pixel-wise metrics, it is indeed low due to the information lost within the VAE process.