MOSS-Image Turbo: 8-step and 4-step versions
The official model wants 50 steps with CFG, which is 100 passes through the transformer per image. Two community distillations cut that to 8 or 4 passes. OpenMOSS says its own few-step distillation (DMD) is in the works.
| Option | Steps / CFG | How to load | H100, transformer only |
|---|---|---|---|
| Official SFT | 50 / 3 | — | 16.2 s |
| lodestones 8-step DMD LoRA | 8 / 1 | LoraLoaderModelOnly after the SFT model: lora.safetensors at 1.0, or lora_latest.safetensors at 0.7 | 1.54 s |
| SFT Turbo INT8 (merged) | 8 / 1, euler, simple, shift 5 | Load Diffusion Model; no LoRA, no negative prompt | — |
| rapid-moss 4-step adapter | 4 / none | Its own nodes or its native workflow rapid-moss-4step-native.json | 0.81 s |
| DeBand Fast INT8 | 10 | Load Diffusion Model; banding fix plus the 8-step LoRA at 0.35 | — |
H100 times from the rapid-moss model card. On an RTX 3060 12 GB, one tester went from 137 s to about 24 s per 1-megapixel image with the 8-step version.
Setting up the 8-step LoRA
Start from the native SFT template (see ComfyUI setup).
Put lora.safetensors (1.19 GB) in
models/loras.Add LoraLoaderModelOnly between Load Diffusion Model and the sampler, strength 1.0. Its author confirmed in the 3060 thread that loading it as-is in the stock loader works.
Set the KSampler to 8 steps and CFG 1. Remove skip-layer guidance: it only acts on the negative branch, and at CFG 1 there is none.
The card also describes an older route with an fp32 transformer and a custom data-space sampler. The author marked it deprecated; ignore it unless you use the Python code.
What you give up
From the 35-prompt test on an RTX 3060: the 8-step version stays close to the full model on text and layout boxes, but drifts on lighting (“chiaroscuro” came out evenly lit) and fine constraints like clock times, a chess position or a fish-eye lens. With no CFG, short prompts look generic and a negative prompt does nothing. Both versions still fail negation (“no cars”) and draw clocks at 10:10.
The rapid-moss card reports more detail and livelier compositions at 4 steps than the 8-step LoRA run at 4 steps, with some regressions shown on the card. MATLOWAI's turbo presets warn to stay near 1 megapixel, since bigger sizes start to tile.
Try it first
The 8-step DMD Space runs the distilled model in the browser. Compare it with the full model in the official demo on the same prompt.