It remains an open question how far it can be pushed. Karras et al have avoided a lot of enhancements they could've used (self-attention being the biggest one)

It remains an open question how far it can be pushed. Karras et al have avoided a lot of enhancements they could've used (self-attention being the biggest one) and don't even train that much (8 GPUs is hardly anything these days), and if you look at BigGAN samples on ImageNet, it manages 'entire scenes, or multiple entire-body photos from different angles' far better than StyleGAN 2 on CATS. (It doesn't do faces well, but what do you expect? ImageNet isn't face-focused and faces make up only a tiny fraction of its pixels.) So, since S2 can do such great faces and BigGAN can do great entire scenes, it seems beyond obvious that it is possible to do both simultaneously and it is 'just' a matter of compiling enough data and training on TPU pods long enough. We hope to run an experiment with https://www.gwern.net/Danbooru2019 https://www.gwern.net/Danbooru2019 and our TFRC TPUs very soon to test the question of how well S2 can scale up...