About the performance on Rain100L dataset in single-task setting.

[csguoh]

Hi, authors. Thanks for your awesome work! I have a question about the PSNR on Derain task under single-task setting. In the original Restormer paper, the Table 1 in this paper show the PSNR on Rrain100L is 38.99dB. However, in the PromptIR paper, the Table 3 show the PSNR of Restormer on Rain100L is 36.74dB. I have no idea about this. Is there some setting different between the two? Can you give a further explanation?

Thank you!

[va1shn9v]

Hello, Thanks a lot for taking interest in our work. Restormer paper and PromptIR paper use different datasets for training on deraining task. Restormer uses a combined dataset of around 14000 images, while we use Rain100L, which consists of 200 large images

[csguoh]

Thanks for your reply. By the way, i have another question that the Prompt technique is usually used for tuning a pre-trained model to adapt to other tasks. But it seems that PromptIR makes all the parameters trainable. Could the authors give me some insight about this design?

Thanks :D

[chensming]

Thanks for your reply. By the way, i have another question that the Prompt technique is usually used for tuning a pre-trained model to adapt to other tasks. But it seems that PromptIR makes all the parameters trainable. Could the authors give me some insight about this design?

Thanks :D

Hi, bro, I have the same doubts. Have you figured it out yet?