A rank 128 LoRA trained with Distribution Matching Distillation (DMD). When loaded onto MOSS-Image-Base, it generates images in 8 steps with CFG 1 and no negative prompts, replacing the base model's 50-step guided sampling. The authors state that it also works with most MOSS fine-tuned versions and other LoRAs, with strength set to 1. The repository provides ComfyUI LoRAs, Diffusers PEFT adapters, example images with embedded workflows, and ComfyUI workflow JSON files. The license follows the base model's Apache 2.0.
Why it matters The MOSS base model defaults to dozens of steps for guided sampling; this LoRA compresses image generation to 8 steps. The authors claim it can be stacked on most fine-tuned versions and other LoRAs.
Hugging Face likes 7 · Hugging Face trending score 7 · Hugging Face downloads 0
License: Commercial use allowed · ComfyUI: Mentioned in model card · VRAM: Not specified
Reasoning
- New text-to-image model on Hugging Face
- ComfyUI: The model card describes ComfyUI usage; the same series / base model already has native support (v0.43.0: MOSS-Image edit LoRAs), but this repository does not confirm it separately
- Score 37 (freshness 15 / source 5 / heat 0 / comfy 4 / vram 3 / demo 0 / license 10)
Usage
- ComfyUI: Drag the 8-step workflow from the example images or workflows into the canvas; place the LoRA in models/loras and load it using Load LoRA (Bypass, Model Only), with sampler lcm, scheduler simple, and CFG 1
- To change shift, use the ModelSamplingAuraFlow or ModelSamplingFlux nodes; do not use ModelSamplingSD3 (the authors state this causes noise)
Risks
- The authors state it was trained for only one sampler; switching samplers degrades performance. Using a standard LoRA loader merges the changes into bf16 weights, resulting in most of them being lost.
- Only 1 example images are provided, and no comparison with the base model's 50 steps is given.