lodestones

/

Chroma1-Radiance

like 110

From Lodestones' post:

Ideally, we want a model that works directly with pixels, without compressing them into a latent space where information gets lost. Ever notice messed-up eyes or blurry details in an image? That's often the VAE hallucinating details because the original high-frequency information never made it into the latent space.

This is the whole motivation behind Chroma1-Radiance. It's an end-to-end model that operates directly in pixel space. And the neat thing about this is that it's designed to have the same computational cost as a latent space model! Based on the approach from the PixNerd paper, I've modified Chroma to work directly on pixels, aiming for the best of both worlds: full detail fidelity without the extra overhead. Still training for now but you can play around with it.

I guess it's either not compatible with Comfy yet, or requires a different workflow.

Thanks for the reply. Is there a node you could recommend for loading into a workflow in ComfiUI? Then I could probably get an output image by experimenting.

You need to be running Comfy from https://github.com/comfyanonymous/ComfyUI/pull/9682 to try this out at the moment. Normal Comfy won't work with it.

I will wait for support in the standard comfyUI, because I don’t understand what I should do to make everything work

HOW TO RUN IT ON FORGE UI?

HOW TO RUN IT ON FORGE UI?

AFAIK only the Chromaforge fork can run Radiance
https://github.com/maybleMyers/chromaforge
Memeory management might be messed up too, hits 24GB VRAM even at 1024X1024

I will wait for support in the standard comfyUI, because I don’t understand what I should do to make everything work

@Viennar Support has been merged into ComfyUI