Bijutsubu (art-club) Anime Art Model

Base: EveryLoRA v1 - a Pony/Animagine/etc hybrid (does not use pony_ tags).

LoRA compatability: Works with most SDXL and Pony-based LoRAs.

Training: Trained using 10 x A40s on a subset of ~390,000 images from the KBlueLeaf/danbooru2023-webp-4Mpixel dataset. 4 epochs + testing steps.

Data selection: Selected over 4,000 most popular artists (by average score on 100 most recent uploads), and sampled up to 100 of their most popular artworks.

Merging: This model will merge with almost anything (including Pony models) thanks to the EveryLora TE block. If you merge with Pony you will transfer over the characters and styles of this model, but will lose the score_* tags. If you do make merges or fine-tunes of this model, I kindly ask that you credit me by linking back to this model.

Credit: Thank you to the creators of Pony, Animagine and SDXL for their models. Thank you to Kohya for their training script. Thank you to Kohaku for their dataset. Big thanks to KafyaHaarukku for their training notebook and settings.

HuggingFace: advokat/bijutsubu

License: https://freedevproject.org/faipl-1.0-sd/

Typical image output (including imperfections):