A StyleGAN2 model (256px resolution) trained on Danbooru2019, with 2x feature maps and noconst finetuning.