Booru Madness

To use:
This is the diffusers version. Put diffusion_pytorch_model.safetensors in your /path/to/pixart/diffusers/transformers folder.

Trained on a random grab bag of over 80k fully tagged 1boy/1girl images + some wikiart. It's got some furry, it's got some monstergirls, it's got some regular anime. How much of which, who knows I didn't count them, it was literally random.

More or less understands a lot of booru tags, mostly unfiltered. It's undertrained, it's a proof of concept of what someone with a 4090 and a week(ish) can do. Some quirks: highly rated images were given the "best_quality" tag, somewhat highly rated images have "good_quality". Wikiart images all had "fine_art" as a tag as well as a style tag ("style_impressionism" for example) as well as the artist name tagged "artist_wikiartname".

It understands booru tags and it understands English sentences although it wasn't trained on sentences, only generated wdv3 tags or the tags as found on the booru sites.

It's trained at 1024px using their recommended multiscale resolution settings, mostly using CAME but some Sophia but Sophia proved unstable and unreliable.

Licensed with whatever Pixart requires.