Joonghyuk Shin

There's little bit of problem in saving fake images with current master branch. I've tweaked few codes to make -sf option work temporarily. You can use the branch in [https://github.com/alex4727/PyTorch-StudioGAN](url)...

I think there's a bug regarding -v option. For now, instead of saving it as a canvas (its what -v option does) you can try to save images one by...

@lavish619 Sorry for late reply, You are correct, wherever mixed-precision is used, an additional condition of `not is_stylegan` is present. That is because StyleGAN incorporates fp16 datatypes in the model...

Hi does this problem still exist?!

I have double checked the code and weight with latest settings and didn't face any problem. It seems like you have used checkpoint for StyleGAN2-ADA. Did you choose the proper...

Hi, StudioGAN **does not provide** implementation nor pre-trained weights for **diffusion models**. To minimize miscellaneous errors in implementation, we have utilized official pre-trained weights and codes for existing diffusion models....

You can use the exact command that you've used when training with a single addition of `-ckpt LOCATION_TO_YOUR_CHECKPOINT_FOLDER` option. The model will then train from the last saved checkpoint. If...

Sorry for late reply, 1. Correct, if you are using a custom dataset it should be pointing to your custom yaml file. 2. No, they do not have to be...

Hi, FID and IS evaluation codes are mostly composed of GPU (CUDA) codes so they work faster. PRDC on the other hand, tends to take longer time as it runs...

I have checked that this issue arrises when PyTorch version 1.12 (the one on the latest docker image) is used and fixed it two days ago! Make sure that you...