MOSS-Image-SFT-8-steps-dmd

MOSS-Image-SFT-8-steps-dmd

Cuts MOSS-Image-SFT image generation down to 8 steps with no CFG, shipped as LoRA files for ComfyUI. A work-in-progress distillation for people who generate images locally.

Made by lodestones on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MOSS-Image-SFT-8-steps-dmd is a distilled version of the MOSS-Image-SFT image model. The original needs 50 steps with CFG 3, and this one needs 8 network evaluations with no CFG.

It uses data-space DMD, where the generator refines its own clean estimate starting from fixed noise. It was trained on compact JSON scene_graph captions, and the card calls it work in progress.

Who it suits: People who generate images with MOSS-Image-SFT and want fewer steps per image, plus researchers who study few-step diffusion distillation.

Industry | Who uses it | How they use it | Example
Design | Image creators who use ComfyUI | Load the LoRA to render in fewer steps | Generate with 8 steps instead of 50
Software | Developers building image features | Cut the network evaluations per image | Sample an image with no CFG
Research | Diffusion researchers | Study data-space DMD distillation | Compare the 8-step output with the 50-step model

Skip it if you want a finished, stable release, since the card marks it as work in progress.

Where to use it:

• ComfyUI, through its LoRA loader
• Windows, macOS or Linux with Python, to download the files
• A demo Space on Hugging Face, which the card points to for a quick look

Use case | Who | What they get back
Faster image generation | ComfyUI users | Images from 8 network evaluations with no CFG
Quality comparison | Researchers | A way to compare against the 50-step original
Experiments with fewer steps | Tinkerers | Output from runs shorter than 8 steps

How to set it up:

• Python and pip to run the download commands
• ComfyUI with the MOSS-Image-SFT base model
• Enough disk space, since the repository holds float32 weights of about 20.3 GB by Hugging Face estimate

1. Install the Hugging Face tools.
2. Download the files from this repository.
3. Add a LoRA loader node in ComfyUI and select a LoRA file.
4. Set its strength to the value the card recommends.

pip install -U huggingface_hub
hf download lodestones/MOSS-Image-SFT-8-steps-dmd

How to use it: Load one of the two LoRA files in your ComfyUI LoRA loader, then sample with 8 steps and no CFG. Each file has its own recommended strength.

File | What it does
lora.safetensors | LoRA to load at strength 1
lora_latest.safetensors | Newer LoRA to load at strength 0.7

Example: You add a LoRA loader to your MOSS-Image-SFT workflow, pick lora.safetensors at strength 1, set 8 steps and turn CFG off. The sampler then produces the image in 8 network evaluations.

For newcomers and experts: If you are new to ComfyUI, start from a MOSS-Image-SFT workflow that already works, add the LoRA loader and change only the strength and step count.

If you know the field, this is data-space DMD distillation of MOSS-Image-SFT. An older section of the card says fewer than 8 steps also work, because each prefix of the schedule is trained. It also warns that in the full-weight route the distillation delta is too small for bf16.

Start from the MOSS-Image-SFT base
Load the LoRA at its recommended strength
Sample with 8 steps and no CFG
Get the image

Good to know:

• The card calls it work in progress and marks one section as deprecated
• The checkpoint is update 2600 with EMA, in float32 weights
• A demo Space exists on Hugging Face

Title: Steps per image
MOSS-Image-SFT original: 50 steps
This distillation: 8 steps

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the files you download and the prompts you give your image workflow
• Not for: the source states no restrictions

On Hugging Face, 13 November 2026

• Library: diffusers
• Parameters: 5.1B
• Weights: about 20.3 GB in F32; about 2.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 333,446
• Likes: 258
• License: apache-2.0
• Based on: OpenMOSS-Team/MOSS-Image-SFT
• Quantized versions: 1
• Tags: diffusers, safetensors

Links

• Model: https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd
• Files: https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download lodestones/MOSS-Image-SFT-8-steps-dmd

Then follow the usage in its model card. Nothing in this kit runs by itself.