Experiments in using BigGAN to generate anime faces and whole anime images; semi-successful.
Release of a 256px BigGAN model trained on Danbooru2019 & e621. This is a prototype model testing our ability to train a BigGAN stably for hundreds of thousands of iterations on a TPU-256 pod on 3 million+ anime/illustration images. While the generated samples are far from ‘photorealistic’, they serve as proof of concept that—unlike our failed StyleGAN-2 scaling experiments—BigGAN can successfully model anime images with great generality, and that we can potentially scale up to 512px or even 1024px and match the DeepMind ImageNet BigGAN for quality.