A hybrid model combining FLEX.1 alpha and T5XXL Unchained, trained on ~1 million curated anime-style images from Danbooru (1 epoch).
Prompt Format Example:
Artwork by [artist], 1girl, rating: {general|sensitive|questionable}, [character details]... (Danbooru tag)
[Natural language description]...
Known Issues
Anatomy inconsistencies in complex poses
Rating tags (general/sensitive/questionable) show limited effectiveness
Requirements
T5XXL tokenizer patch MUST be applied (https://github.com/Kaoru8/T5XXL-Unchained)
CFG is needed for this model, the distilled module was broken during training. Recommended CFG Scale is around 4.
Updated versions may be released pending continued development.
Description
Experimental checkpoint trained based on modified models.