Hugging Face's logo

KBlueLeaf
/
Kohaku-XL-Zeta

Text-to-Image
Diffusers
Safetensors
English
StableDiffusionXLPipeline
Model card Files
xet
 Community
3

Anima fine-tune

#3
by kdutt2000 - opened
Discussion
kdutt2000

I'm a huge fan of your work and all the work you've done over the years.

Would ever consider doing a fine-tune of a great open source anime model co developed by Comfy Org called circlestone-labs/Anima.

The creator tdrussell is really good as well and each preview version has been a big improvement over the other. It would be awesome if you could collaborate with him too.

They now have a official turbo LORA too.

KBlueLeaf
 Owner
•
edited May 7

Just FYI I'm also employee of Comfy Org and already working there for year (part-time tho) (note I'm not engaged in Anima)

I do plan to do more stuff about anime t2i but now more focus on MLLM stuff (as I'm doing research related to VQ/token AR image gen)

And I may also investigate the possibility to continue the SDXL-neo project which aiming to improve SDXL to be more modern and on par with DiT based stuff.

But as I'm more focusing in research or some other opensource stuff (like KohakuTerrarium agent framework) I can't gurantee when will I obtain new model here.
KBlueLeaf
 Owner

Few "project" that I may start:

  1. Kohaku-DiT (directly pretrained from scratch, and larger than HDM)
  2. Kohaku-MLLM (native multi modal AR token gen model, but more focus on media generation)
    • can be fully trained from scratch or tuned from open sourced model
  3. SDXL-neo
    • I may consider to finetune current SDXL with following tech:
      • flow-matching + logitnorm sampling
      • Better VAE (with my latentmaird work, paper/code release soon)
      • Better TE (directly replace with better TE like some small LLM or use ELLA like adapter approach)
    • Will be general t2i at first, may trained with few dozen M general + anime images.

Sign up or log in to comment