Open Ding3LI opened 2 months ago
Hello,
Everything you are doing seems fine, it's just that in order to use StyleGAN3 you will need to modify the style injection in the same way I did for StyleGAN2. If you diff networks.py
between this repo and the original stylegan2-ada code, you will be able to see what changes were made. This can be a bit tricky but in principle I don't see why it wouldn't work. If you are changing network architecture you may also need to update the learning rate to get proper results.
Let me know if you try doing this & need help!
Hi, thanks for sharing such great repo for solving image degradation. I am currently trying to update the code from styleGAN2-ADA to styleGAN3. I simply modified some imports since styleGAN3 uses different name and path, and what I understand is that this image restoration method is mainly doing image generation (please correct me if I misunderstood). I pasted below:
So, I downloaded a new pre-trained model from styleGAN3 named
stylegan3-r-ffhq-1024x1024.pkl
. I keep using the same datasets that this repo provided (sample_1.png and sample_2.png), and all other codes remain same. \ However, the program can run forW
andWp
successfully, but it will be terminated by an error when enteringWpp
process. The error message is copied below:Could you share any suggestions on solving this issue? Thanks.
My venv:
TLDR
This environment can run both styleGAN2-ADA and styleGAN3.