search

yangxy / GPEN

2.39k stars 449 forks source link

About reproduce face restoration with pretrained StyleGAN2 Model #122

Open AbnerVictor opened 2 years ago

AbnerVictor commented 2 years ago

Hi, thanks for sharing the great work. I'm interested in reproducing the GPEN-BFR-512 from zero, since it takes too much time to reproduce 1024 version, however I got some issues

  1. I wonder which pretrained StyleGAN2 checkpoint you adopt. It seems that it's not provided.
  2. Did you use pretrained discriminator as well?
  3. The results look blurry in first 20K iterations for 1024 reproduce training, is it reasonable? image
yangxy commented 2 years ago
  1. You should train it from scratch as it is not publicly available.
  2. Yes.
  3. The result looks terrible and not right from my perspective. Please check if the outputs during training are plausible.
ketyi commented 11 months ago

Hi @AbnerVictor, Were you able to obtain a pre-trained StyleGAN2 checkpoint on FFHQ trained on 256^2 resolution?

zimenglan-sysu-512 commented 10 months ago

+1, @AbnerVictor can u share the information about how to pretrain the StyleGAN2