Finetuning a StyleGAN2 FFHQ model to a collection of pixel characters and purposely underfitting it sure leads to some interesting results.

I finetuned NVIDIA’s FFHQ256 Stylegan2-ada-pytorch model by showing it only about 20.000 instances of 64x64 pixel characters (upsampled to 256x256). This took around 30 minutes on Colab. Here are some more (uncurated) results.

(download a high-res image here)