MOSS-Image 8-step DMD

MOSS-Image 8-step DMD

Generates images from text in 8 steps with a distilled version of MOSS-Image-SFT, in a Gradio page on ZeroGPU hardware. A work-in-progress demo for people who want faster text-to-image.

Made by hugging-apps on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: MOSS-Image 8-step DMD is a text-to-image demo for a work-in-progress distilled copy of MOSS-Image-SFT, published as lodestones/MOSS-Image-SFT-8-steps-dmd. Where the base model spends 100 network evaluations on one picture, this one spends 8 and skips CFG. It runs on ZeroGPU hardware, so you can try it in a browser.

Who it suits: It suits creators, researchers and developers who want to see how much a distilled image model gives up for the extra speed.

Industry | Who uses it | How they use it | Example
Content creation | Creators | Get quicker pictures from a text prompt | Type a short prompt and wait for the result
AI research | Researchers | Study a distillation of MOSS-Image-SFT | Try the same prompt at different LoRA strengths
Software | Developers | Read how the pipeline applies an unmerged LoRA | Open app.py and pipeline_moss_image.py

If you need the full quality of the base model, you can skip it, since the README calls this a work in progress.

Where to use it:

• A web browser, on the Space's page on Hugging Face
• A local copy of the files, run on Python 3.12 under Windows, macOS or Linux

Use case | Who | What they get back
Faster image drafts | Creators | A picture from 8 steps instead of 50
Tune the strength | Researchers | Results at different LoRA strengths
Study the pipeline | Developers | A short pipeline file that loads the base model and the adapter

How to set it up:

• A web browser, which is all the online page needs
• Python 3.12 and the packages in requirements.txt, for a local copy

1. Open the Space's page on Hugging Face to use it online. Nothing needs installing.
2. For a local copy, install the Hugging Face hub package and download the Space's files.
3. Move into the downloaded folder.
4. Install Gradio 6.30.0, then the packages from requirements.txt.
5. Start the app with python app.py.

pip install -U huggingface_hub
hf download hugging-apps/moss-image-8-step-dmd --repo-type space --local-dir moss-image-8-step-dmd
cd moss-image-8-step-dmd
pip install gradio==6.30.0
pip install -r requirements.txt
python app.py

How to use it: Open the page, or the one your local copy starts, write a prompt and generate. Ordinary sentences and tag lists work as they do in the base model, even though training used scene-graph captions written as compact JSON.

Tool | What it does
Prompt | Takes a natural-language sentence or a tag sequence
LoRA strength | Sets how strongly the distilled adapter is applied, with 0.7 as the starting value

Example: Type a short description of a scene and generate with the default strength. The page returns an image made in 8 steps. If it looks too saturated, lower the strength toward 0.6 and try again.

For newcomers and experts: Open the Space page and run one short prompt at the default strength. On PartiPrompts, strengths from 0.6 to 0.8 followed prompts best. At 1.0 the colors run a little too saturated, and under about 0.5 the 8 steps cannot do enough.

The distilled LoRA has rank 128 and is applied unmerged on top of the stock MOSS-Image-SFT transformer, with the Qwen3.5 text encoder, the tokenizer and the FLUX.2 VAE from the base repository. It samples with the loop the base model already uses, an Euler flow-matching sampler at shift 5, for 8 steps and without CFG. Two extra parts of the distillation, a step embedding and a latent mix, are not used, since they made short prompts come out as grids with invented poster text.

You write a prompt
The base model's text encoder reads it
The unmerged 8-step LoRA is applied at your strength
An 8-step Euler loop makes the latents
The VAE turns them into the picture

Good to know: The README calls the distillation a work in progress, so results can differ from the base model.

• License: Apache 2.0 for the model weights, and BSD 3-Clause for the bundled pipeline code
• Cost to run: none named, and the online copy runs on ZeroGPU hardware
• Data it reaches: the prompts you type and the models the Space names
• Not for: work that needs the base model's full 50-step quality

On Hugging Face, 13 November 2026

• Likes: 34
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/hugging-apps/moss-image-8-step-dmd
• Files: https://huggingface.co/spaces/hugging-apps/moss-image-8-step-dmd/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.