I saw the section about not training from scratch (via transfer learning) in https://www.gwern.net/Faces#transfer-learning. The Holo example is really impressive!

How expensive is it in terms of labelled data and compute? Do you know if anyone tried this for just ahegao faces?

> How expensive is it in terms of labelled data and compute?

All the stuff you see in that page (except the BigGAN ones) is unconditional, no labels. You just dump the images in and it figures it out. StyleGAN does support labels via one-hot embedding as I understand it, but I don't know how to use it so none of my experiments use it. A few people have mentioned or used it, but there's no good documentation about how to make it work, so... For unconditional samples, it depends on how many you have and how different they are. You can see in the various examples transfer learning with a few hundred to a few thousand (with and without data augmentation).

> Do you know if anyone tried this for just ahegao faces?

It's funny you ask that because I was corresponding with an anonymous who was using it for just that (and ball gags). He'd run into some issues with the encoder/editing functionality and wanted advice, but the regular transfer learning worked fine. He'd compiled a small dataset of a few hundred to a few thousand examples on his own, and it worked disturbingly well: sufficiently so I didn't want to write it up. (I try to keep my site SFW.)

Thanks! I was just in the process of digging through your stuff.