catlab-team / latentclr Goto Github PK

Hi there,

Thanks for the interesting work. Are you planning on sharing pre-trained weights for the models?

If so, I'd be particularly interested in the weights for StyleGAN2 on FFHQ (all the types of direction models would be great).

Than you!

Hi,

This work is interesting!
Did you release the code for evaluation metrics like Rescoring Analysis?

Thanks!

Hi,
I think your idea that discovering directions using contrastive learning[https://arxiv.org/pdf/2104.00820.pdf] is almost the same with "Do Generative Models Know Disentanglement? Contrastive Learning is All You Need" https://arxiv.org/abs/2102.10543 on arxiv, which was submitted into arxiv two months ago.
I think you should talk about it seriously.

Hi, thanks for sharing your nice paper. In your paper, you mention that the code is publicly available but the repository is empty. Is this April fool or something?

Hi,

Currently I'm using StyleGAN2 and extract the output from feature layer "conv1" as mentioned in the paper.
The dimension of the input layer in NonlinearProjector, according to conf/size/stylegan2_nonliner.yaml, is 512. But I got 8192 instead since the output size from 'conv1' is actually [batch, 512, 4, 4]. Is it because I set use_w to True or I should overwrite the dimension setting in stylegan2_nonliner.yaml?

Thank you