LORA by reakaakasky

Enhancement bundle for Chroma HD

Cover images are raw output from the model, 1MP resolution, no upscale, no hands/faces inpainting fixes, even no negative prompt. They are not perfect, but 100% "what you see is what you get".

This model

This is an all-in-one (all aspects I can come up with) finetuned model.

Trained with total ~5k images (latest or recent versions). Mixed with photograph and anime images.

• No AI images. Not on my watch. Every image is hand-picked by me. Only high resolution images with finest details. The whole dataset avg pixels is 3.37 MP, ~1800x1800.
• Comprehensive natural language captions from Google LLM.
• All anime characters are tagged by wd tagger v3 first and then refined to natural language by Google LLM.

Effects:

For general realistic image:

• It gives you better natural texture and details.
• Better lighting, a little bit cinematic feelings.
• Less overexposure and oversaturation.

For anime:

• Because Chroma does not support artist tags, and quality tags (not sure though, in my tests, the base model didn't response to those tags.). It's very unstable for anime.
• I added a default anime style. This default anime style is mainly for reducing noise and overfitted bias that towards simple low quality image.
• Also included a sub dataset full of hands.

How to use:

Simple, this is a LoRA, apply it with strength 1 then you will get the finetuned model.

For anime, use "Digital anime illustration of ..." as trigger words.

Recommended settings: I use euler_cfg_pp with cfg 0.5~0.7, beta schedular, 30 steps.

License