Trained on HDTF dataset for 10k steps, droplet artifacts found

[veryverypro]

@lucidrains I've trained upsampler on HDTF dataset for 10k steps, and the sample image shows lots of droplet artifacts that suppose to be fixed by StyleGAN2 mod-demod operation, instead of AdaIN. Why?

And I've checked your code with official stylegan2 implementation, I notice that you init all conv layer weights with kaiming init, which does not exist in the original implementation of stylegan2. Could that be the problem?

Looking forward to your reply. Thanks in advance.