Introduction

MOSS-Image just got a lot faster. lodestones, the trainer behind the Chroma image models, has released two Apache-2.0 distillation LoRAs that take OpenMOSS's 5B text-to-image model from 50 steps with CFG 3 down to 8 steps with no CFG. The SFT version went up on Hugging Face on October 16 and has already passed 12,000 downloads, and a Base version followed this morning, October 18 (about 01:30 UTC). Both drop straight into a standard ComfyUI LoRA loader.

We covered MOSS-Image at launch and when it went native in ComfyUI. This is the speed follow-up: same model, same layout tricks, roughly a twelfth of the compute per image.

What Shipped

There are two repos, one per official checkpoint:

  • MOSS-Image-SFT-8-steps-dmd: distilled from MOSS-Image-SFT, the fine-tune with tighter text grouping. The checkpoint is update 2600 (EMA). There's also a free Hugging Face demo Space running it.
  • MOSS-Image-Base-8-steps-dmd: distilled from MOSS-Image-Base, the general-purpose generator. The checkpoint is update 9200 (EMA). It's brand new, with 5 likes at the time of writing.

Both model cards call these a work in progress, so expect new checkpoints. A few more details from the cards:

  • The LoRA is rank 128 / alpha 128 across every block's attention and feed-forward projections, stored in fp32, about 1.19 GB per file
  • It uses kohya-style keys, which is why ComfyUI's stock LoRA loader picks it up
  • The license is Apache-2.0, same as MOSS-Image itself
  • Both were trained on compact JSON "scene_graph" captions, which suits MOSS-Image's JSON layout prompting

Why 8 steps with no CFG is a big deal

The teacher runs 50 steps at CFG 3, and classifier-free guidance costs two passes per step, so that's about 100 forward passes per image. The distilled LoRA needs 8 passes, roughly 12 times fewer. On a mid-range card that's the difference between waiting on every render and actually iterating on a layout.

The method: data-space DMD

The cards cite "Few-Step Generation via Data-Space Iteration" (arXiv 2610.12102, October 8), by Shanchuan Lin, Yansong Peng, Fu-Yun Wang and Haoqi Fan. Instead of stepping along a fixed noise schedule, the generator starts from noise and keeps refining its own clean-image estimate, and the paper plugs this into distribution matching distillation (DMD). One practical upshot, per the cards: every prefix of the 8 steps is trained, so fewer steps also work.

How to Run It in ComfyUI

You need a working MOSS-Image setup first. The native ComfyUI support and the Comfy-Org/MOSS-Image weights from our earlier guide are the easiest route. Then grab the LoRA for the checkpoint you use:

# SFT distill (pair with the MOSS-Image SFT diffusion model)
hf download lodestones/MOSS-Image-SFT-8-steps-dmd lora.safetensors --local-dir ComfyUI/models/loras/moss-sft-dmd

# Base distill (pair with the MOSS-Image Base diffusion model)
hf download lodestones/MOSS-Image-Base-8-steps-dmd lora.safetensors --local-dir ComfyUI/models/loras/moss-base-dmd

Put a LoraLoaderModelOnly node between your MOSS-Image diffusion model and the sampler, and match the LoRA to the checkpoint.

Settings from the model cards

  • Base LoRA: strength 1, the model's usual flow-matching Euler sampler, shift 5, CFG 1, and 4 to 8 steps. The card says nothing else from the repo is needed.
  • SFT LoRA: lora.safetensors at strength 1, or the newer lora_latest.safetensors at strength 0.7. The card describes it as 8 network evaluations with no CFG, so the Base settings above are a sensible starting point.
  • Keep CFG at 1. The distill has guidance baked in, and turning CFG back up just doubles your passes.

The repos also include a raw fp32 transformer and a custom "data-space" sampler snippet, which is the training-time form. The SFT card now marks that path as deprecated, and the Base card recommends the LoRA instead, so most people can ignore it. If you do go that route, the cards warn that the distillation delta is smaller than bf16 rounding, so that transformer has to load and run in fp32.

A prompt to test it

Try something with rendered text, since that's MOSS-Image's party trick:

A vintage travel poster of a lighthouse on a rocky coast at dusk, lilac and teal sky, warm gold beam sweeping across the sea, bold serif headline "NORTH HARBOR" across the top, small caption "Ferries daily" at the bottom, flat screen-print texture

Run it once on the full 50-step setup and once on the 8-step LoRA, and compare the text and the fine detail. That tells you quickly whether the distill holds up for your kind of work.

Original Source

Model cards: MOSS-Image-SFT-8-steps-dmd and MOSS-Image-Base-8-steps-dmd. Demo: MOSS-Image 8-step DMD Space. Paper: Few-Step Generation via Data-Space Iteration.

Conclusion

MOSS-Image already had the layout control and the bilingual text. Now it has the speed too: one LoRA, CFG 1, 8 steps, Apache-2.0. It's still a work in progress, so keep an eye on the repos for newer checkpoints. For poster mockups and thumbnail drafts, though, this is the fastest way to run MOSS-Image locally today.

—Aurelia ♡