Kroma
Converted from https://huggingface.co/lodestones/Kroma/tree/main
All credit belongs to the original model author. License is the same as the original model.
Quantized qwen3vl text encoder: link.
w4a8 xl:
This version is optimized for 12GiB VRAM. The file size is ~9.2 GiB, ~40% layers are int8convrot. This ckpt should be able to fully load to the VRAM without offloading/swapping (Assuming your OS won't allocate more than 1GiB VRAM), and you get the maximum speed and better precision (Similar to GGUF XL suffix).
int8convrot and others:
https://huggingface.co/Stabhappy/kroma-v0.2-base-INT8-convrot
Description
FAQ
Comments (6)
This model is very good at prompt adherence
This is a piece of crap, broken in every way imaginable. I don't know what would have to be done to Crea2 to make her a complete cripple on the level of the odious Pony7. Lodestone is a windbag.
If you have time please can you do a w4a8 format for Anima
What's the point. int4 weight has avg 7% loss.
Is it some very fresh abbreviation of format int8/4 that requires newest comfyui? As it throws me error on attempt to load this checkpoint in both my workflow (unet loader) and your image embedded workflow (kj loader)
Im on comfy 0.30
Or am i dense and this should not be stored in Models/Unet?
Quick google search and yep, seems i need 0.31 minimum