search

eladrich / pixel2style2pixel

Official Implementation for "Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation" (CVPR 2021) presenting the pixel2style2pixel (pSp) framework
https://eladrich.github.io/pixel2style2pixel/
MIT License
3.2k stars 568 forks source link

Images plotted when training for super-resolution task does not look like downsampled #272

Closed catherineyeh closed 2 years ago

catherineyeh commented 2 years ago

Hi all, I've tried to perform the super-resolution task with a trained stylegan2-ada-pytorch as the generator. However when training, the images stored in the checkpoints folder look like this below: sr_1200

I'm not entirely sure if the image under "Inputs" are downsampled correctly when training. Could it be that the transformations/preprocessing steps need to be modified? The training images were greyscale 256x256.

yuval-alaluf commented 2 years ago

Could it be that the transformations/preprocessing steps need to be modified?

I believe that it could be that the BilinearResize assumes the image has three channels and you are giving it a single channel input. You can try converting your grayscale images to RGB (using something like Image.open("image.png").convert("RGB") and if that doesn't work, you could try replacing the BilinearResize with a different downsampling function that behaves correctly.