Experimental lora for Hiryuu from Kantai Collection for use with Flux.1 Dev based modelsThis is by no means finished or easy to use, just publishing to share my current results and thoughtsFor use with Flux.1 Dev models. I've not tested with other checkpointsI recommend using about 5 distilled CFG scale if prompting for her main outfit.Unlike XL and 1.5 you can actually specify for the number of lines on her sleeves and it will (usually) help. It can take a while but iterating on natural language prompts works pretty well.Trained on a set of 12 high quality images of her main outfit and 6 high quality images of other outfits. The main outfits were captioned manually with natural language, and the 6 other images were captioned with natural language by editing <a target="_blank" rel="ugc" href="https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha">Joy Caption pre-alpha</a> outputs. Then, 100 random permutations of caption ordering were generated for each image (so I could shuffle the captions while still caching TE output)Previews were made using <a target="_blank" rel="ugc" href="https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main">Q8 quantitized FLUX.1 dev</a> along with fp8_e4m3fn T5XXL with WebUI Forge. Euler Beta 40 steps. If the details don't really come out right increasing steps often helps (20 is only sufficient for simple scenes I think).It seems to have some amount of bias towards Wa's style as his Hiryuus made up almost a third of the training set. Including "art by Wa" in captions didn't really seem to help differentiate things which is maybe not surprising considering TE is frozen in Flux lora training.

Files with SHA256: f5e67d83b0a9cfbcf9c2bf6f52379c276b2e70a0c6207666fec60a09d86c8360