training decoder during Stage 2

[zhpenli]

Thanks for the wonderful work! I noticed you only train 100 iterations for decoder, while 25k iterations for SD. Is it enough to train decoder for only 100 iterations? BTW, did you use combination of L1 loss, perceptual loss, and GAN loss to train decoder?

Thanks in advance!

[csslc]

Thank you for your attention! Yes, we utilize a combination of L1 loss, perceptual loss (LPIPS), and GAN loss to train the decoder. During training, we observed that optimizing the LPIPS loss significantly enhances the decoder's ability to generate fine details, particularly in the early stages, leading to impressive results. To further involve the GAN loss in refining these details, we extended the total number of iterations in CCSR-v2, resulting in clearer and more refined outputs.