Played around with @gwern's TWDNEv2 model to generate images of Hayasaka Ai! This is after ~9 hours of training (n = 300+). Stopped working on it after a bit, so a bunch of potential improvements. More thoughts here:…

Oct 11, 2020 · 11:56 PM UTC

Some quick thoughts: * i hate colab * pixiv doesn't tag images properly. ended up writing a short script to filter hayasaka out by hair color. * pretty sure there's a LOT of room for improvement; this was trained for very little time on a Colab Notebook
I may end up writing a proper blog post at some point, but for now, I'd like to thank Gwern for his super in-depth guide at Also helped out by the author of this article on CSDN: