stylegan2-encoder-pytorch icon indicating copy to clipboard operation
stylegan2-encoder-pytorch copied to clipboard

Published 20 hours ago • verified publisher icon bryandlee

The Result

Open YunShen66666 opened this issue 4 years ago • 2 comments

Hi The results of inversion seems not good as the paper. Did you know what the qusetion is?

YunShen66666 avatar Nov 07 '20 12:11 YunShen66666

Hi, I only qualitatively compared the results with the paper, and it seems like the encoder performance does not match with the original one while the inversion looks ok.

image

The original implementation used the StyleGAN1 model provided by Nvidia and the encoder was trained with batch size 128, whereas this repo uses the StyleGAN2 model implemented by rosinality and the encoder was trained with batch size 8 on a single GPU.

bryandlee avatar Nov 07 '20 13:11 bryandlee

Thanks for your careful reply! And I have another question : Since I only have a single GTX2080Ti , training 1000000 iters will take 10 days! , which is too long for me. My task is inversion. Can I train 100000 iters and then do inversion?

YunShen66666 avatar Nov 09 '20 09:11 YunShen66666