12h of transfer learning on gwern's faces model (gwern.net/Faces#anime-faces) on my GTX970 with 988 asashio faces (18772 after data augmentation) Still some green blob artefacts, hope they reduce with some more learning.

Apr 1, 2019 · 11:01 AM UTC

1) Some style mixing samples 2) Random samples
This tweet is unavailable
I initially set learning rate at 0.001; after more training (at ~20h) the blobs were reducing. Although I had some other kind of artefact that ruined the face for some. These reduced after ~22h of training at 0.0005. I was limited by minibatch size of 2 due to my GPU.