You can deploy this model and start earning money today!

rocca/
chroma-nunchaku-test

Name: chroma-nunchaku-test
Availability: InStock

Views

No views yet

NOTE: The quality of outputs produced by the models in this repo are not as good as they could be, probably due to bugs in my code. You may need to wait for official Nunchaku support if you want good quality outputs.

EDIT 1: See this thread for notes on choosing the best v38 model.

EDIT 2: See https://huggingface.co/spooknik/CenKreChro-SVDQ for Chroma/Krea merge quants (IIUC, this works with Krea/Flux workflows, not Chroma/FluxMod).

Context: Chroma is an Apache 2.0 de-distill of Flux Schnell by Lodestone Rock (please support him here). If you'd like to see the advantages it has over Flux Schnell, check out the examples in this post - but note that those are made with Chroma v15, and Chroma is now at v29 as of writing. Still, you can see that Chroma has a more "natural" (non-cartoon, non-plastic) style. It can even do "bad" images (e.g. blurry, low-aesthetics) if prompted. I think it's much more verstatile than Schnell, and has much broader knowledge.

As of writing one problem is that it's quite slow. Lodestone is working on this via some few-step experiments, and it looks like we might get official nunchaku support soon.

This repo contains a few experimental Nunchaku models, produced via hacking the DeepCompressor codebase to support a Schnell-compat version of the Chroma weights. It unfortunately contains some "useless" weights to make it architecturally the same as Schnell, at the cost of increasing the model size by about 30%. That's another reason to wait for official Nunchaku support.

The models are based on this quantization script: https://gist.github.com/josephrocca/dd339fc20c18dd48524c57b6e4005486

Usage:

Install FluxMod. This is needed even though Chroma is officially supported in ComfyUI. We need the padding removal nodes, because we're "pretending" that Chroma is a Flux Schnell model.
Use this Chroma workflow, but swap the Chroma Unified Model Loader node for the Nunchaku FLUX DiT Loader node.
If you want batch size >1, you may need to add a Rebatch Latents node after the Empty Latent Image node, like this.
I think Rescale CFG node can mess with it, causing a burn effect, though I'm not sure, since I haven't thoroughly tested it. Try disabling that if it's in your workflow.

rocca/chroma-nunchaku-test

NOTE: The quality of outputs produced by the models in this repo are not as good as they could be, probably due to bugs in my code. You may need to wait for official Nunchaku support if you want good quality outputs.

Usage:

rocca/
chroma-nunchaku-test