Originally posted by silveroxides in his HF repo.<a target="_blank" rel="ugc" href="https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/tree/main/hyper-low-step">https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/tree/main/hyper-low-step</a>For those who don't like the speed or stability of Chroma - use these two LoRAs (dead links, use both models from this post):<a target="_blank" rel="ugc" href="https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/blob/main/hyper-low-step/Hyper-Chroma-low-step-LoRA.safetensors">https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/blob/main/hyper-low-step/Hyper-Chroma-low-step-LoRA.safetensors</a><a target="_blank" rel="ugc" href="https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/blob/main/chroma-unlocked-rescaled_cfg_LoRA-rank_16-fp32.safetensors">https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/blob/main/chroma-unlocked-rescaled_cfg_LoRA-rank_16-fp32.safetensors</a>Together they allow generating at cfg 1 with lower than 10 steps (8 seems to work well), which gives a very nice speed boost. If you want negative prompt to have an effect, use NAG:<a target="_blank" rel="ugc" href="https://github.com/ChenDarYen/ComfyUI-NAG">https://github.com/ChenDarYen/ComfyUI-NAG</a>or NAGuidance node<a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/embedded-docs/blob/main/comfyui_embedded_docs/docs/NAGuidance/en.md">https://github.com/Comfy-Org/embedded-docs/blob/main/comfyui_embedded_docs/docs/NAGuidance/en.md</a>but keep in mind that it will decrease generation speeds.optimal values for Normalized Attention Guidance node:nag_scale=3.0nag_alpha=0.25nag_tau=1.0nag_sigma_end=0.75

Hyper Chroma low step LoRA

Hyper-Chroma-low-step-LoRA.safetensors

Chroma-low-step-LoRA

rescaled_cfg_r16-fp32

This is not my work but from <a target="_blank" rel="ugc" href="https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/tree/main">https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/tree/main</a>The intention is to be able to upload images created with it and added the lora to it.

Files with SHA256: 932689d13dfa4b2cd910b3ce6e4248706a8d056afada125567e874386468ee5f