This LoRA is not mine, The author is lodestones, I'm just reprinting it for general purposes.
Trained with Danbooru art.
https://huggingface.co/lodestones/Kroma
This LoRA was produced from a K2 (Krea 2) fine-tune checkpoint and converted into the ComfyUI format: LoRA adapters use the native diffusion_model.* key naming, and the fully-fine-tuned RMSNorm / modulation tensors are stored as .diff deltas against the base model. Both are applied by ComfyUI's standard LoraLoaderI will delete it if the original author requests it.
Description
FAQ
Comments (9)
what does this lora do?
Looks like it's a general fine-tune https://huggingface.co/lodestones/Kroma
@frauleinpiss1234 still but what does it do? i am noob , new to this.
@opie1 No idea I'm afraid, it doesn't say on the huggingface page
as far as i know it's a large scale tune that was trained on danbooru, e621, art and other general stuff. For me, it improves the prompt adherence and doesnt censor stuff anymore (i only generate anime/illustration tho). It was trained at only 256x256 which Ive noticed in my generations. Im still testing but 0.75 seems like a good strenght
It's Lodestones, the guy that finetuned Chroma, so you can expect it to have a vastly improved understanding of all booru tagged content.
Practically, expect it to know a lot more niche stuff and to better understand what you want with short tags instead of descriptive language.
Important to keep in mind that this is version 0.1, right now it's not very powerful. Chroma went through... nearly 50? versions, and it only really started to be good at version 30. He tends to start small for proof of concept then gradually improve the training.
@lifania tldr another project that will never be finished :)
@deitychaser the lora did not work for me
Has anyone tried version 0.2 (turbo)? Please share your impressions.