StyleGAN2 512px trained on Danbooru2019
GAN-generated anime-style images.
Try out the interactive version: This Anime Does Not Exist (link).
A custom variant StyleGAN2 model (512px resolution) trained on Danbooru2019. Notable changes include disabling stylemixing during training, forgoing path length regularization, using 2x feature maps, noconst (from the start), top-k (at convergence), and data augmentation using PALM (hands and figures augmentations), among other tweaks.
There’s quite a variety of drawing and material styles. Watercolors and ink washes are well represented.
Subjects are diverse, too. Portraits, full characters, scenes.
Some more scenes…
And more portraits…
The model seems particularly enamored of, and adept at, ornamentation. It has a real penchant for painting in floating flowers, stars, and hearts.
There’s also increasingly good coherency, texture, and attention to detail in terms of fashion.
Also: hats. SG2 especially likes large floppy hats.
Some hands are fairly well rendered.
Many… aren’t quite there.
SG2 attempts more unusual poses, with mixed success.
Comic-styles are also somewhat common.
And a few very rare images aspire to contain no human forms at all.
A couple weird ‘dark’ images…
And finally, a few more that stood out.