MOSS Image 8-step DMD LoRA

This is a LoRA adapter published by "lodestones" on Hugging Face. It distills OpenMOSS-Team/MOSS-Image-SFT, which normally uses 50 steps at CFG 3, down to 8 network evaluations with no classifier-free guidance. The method is data-space DMD (distribution matching distillation), in which the generator iterates on its own clean estimate from fixed noise. The model card describes it as a work-in-progress and says it was trained on compact-JSON scene_graph captions.

The adapter is loaded through a ComfyUI LoRA loader. lora.safetensors is recommended at strength 1, and lora_latest.safetensors at strength 0.7. The card also documents a rank-128 fp32 LoRA applied to the attention and FFN projections, plus a custom sampler for use outside ComfyUI, but that section is marked deprecated. Fewer than 8 steps also work, since each prefix is trained.

A YouTube reviewer (AI Search) reports that it makes MOSS Image roughly three to four times faster, at the cost of negative prompts having less effect, because CFG is not used. A demo Space is linked from the model card. The vendor is an independent community author, not the MOSS base-model team.