The first 2D style (Anime) model. Our initial target is to make this model able to be some kind of style "regularization" of LoRA tuning. This is very early version of Haigaku as long as Alpha flag still exists.

This model trained on 1k+ datasets from 2021 Danbooru images including Natural Language captioned only, images are captioned by latest proprietary LLM which has lower censoring level. Trained on 1x Nvidia H100 with 30 batch size, while keeping three text encoders frozen.

Haigaku-Large currently only focusing on style only tuning due to very large transformer parameters, hence the training has slow speed. But, we always encourage to find much more efficient tuning method and increase our capacity to be able produce more concept instead style only.

We recommend using this model with either dual CLIP loader or three CLIP loader TE configuration to achieve best results. This model maybe still producing undesired image like anatomy horror (mostly the culprit is on t5-xxl encoder), or low-quality image. We will improve this model further.

Haigaku-Medium also in development, with broader datasets. We will release it soon