This is not mine. See the original HuggingFace repo here.

This model works exceptionally well with the Rank 32 flash lora. I use Euler/Normal, 24 steps, 1cfg. Yes, 24 steps, but the 1cfg makes it go by very quickly.

I have found this checkpoint to be far more stable than the original base checkpoint: less body morphing, less 6 fingers, less incoherent and broken results.

As you can see from my many examples in this gallery, it can easily do realistic, anime, cartoon, horror, etc.

Use normal Chroma workflow from ComfyUI. Don't both with fancy samplers, schedulers, sigma shifts, etc. Euler/Normal, 24 steps, 1cfg with the flash lora. That's it.