MOSS-Image-SFT-4-steps

MOSS-Image-SFT-4-steps

Makes the MOSS-Image-SFT text to image model finish a picture in 4 steps with no guidance, using one LoRA file. Renders English and Chinese text, for ComfyUI and diffusers users.

Made by akhaliq 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-4-steps is a distilled adapter for MOSS-Image-SFT, a text to image model with 5.1B parameters. It lets the model draw an image in 4 steps, with no classifier-free guidance, using one LoRA file and the standard flow-matching Euler sampler. The repository also holds the pipeline, text encoder and VAE.

Who it suits: Image makers and developers who want fast generations from MOSS-Image, especially pictures that contain readable English or Chinese text.

Industry | Who uses it | How they use it | Example
Design | Designers | Draft signs, menus and posters with text in the picture | A cafe menu board with prices
Marketing | Content teams | Produce quick visuals in English and Chinese | A storefront image with Chinese lettering
Software | Developers | Add fast image generation to an app through diffusers | A script that renders a prompt in four steps
Research | ML researchers | Study few-step distillation of an image model | A comparison of 3-step and 4-step results

If you do not use MOSS-Image models, or want the slower full-quality sampling, you can skip it.

Where to use it:

• ComfyUI, through its LoraLoaderModelOnly node
• Python with diffusers, through the apply_lora.py script in the repository
• Windows, macOS or Linux with Python, for downloading the files
• A browser, through the demo Space the card links

Use case | Who | What they get back
Fast text to image | Anyone with a GPU | A finished picture from a prompt in 4 steps
Bilingual text in images | Designers | Signs and menus with legible English and Chinese
Large outputs | Artists | Images up to 2048 by 2048 with no tiling, per the card

How to set it up:

• Python, to run the Hugging Face download tool
• A GPU, where the card's timing comes from one A100
• ComfyUI, or diffusers with the repository's own pipeline
• About 20.3 GB for the F32 weights, or about 2.5 GB at 4-bit as an estimate

1. Install the Hugging Face hub tools with pip.
2. Download the repository with the hf command.
3. Load the LoRA file in your tool, with LoraLoaderModelOnly in ComfyUI or apply_lora.py in diffusers.
4. Set 4 steps, a guidance scale of 1 and the stock Euler sampler.

pip install -U huggingface_hub
hf download akhaliq/MOSS-Image-SFT-4-steps

How to use it: Write a prompt and generate with 4 steps and no guidance. The card says there is no negative prompt either. It also ships an inference.py command line script.

Tool | What it does
lora.safetensors | The single adapter file that gives the 4-step behavior
apply_lora.py | Applies the adapter in diffusers
LoraLoaderModelOnly | Loads the adapter in ComfyUI
inference.py | Command line script for generating images

Example: You ask for a cafe menu board that reads "TODAY'S SPECIAL, Flat White $4.50". The model draws the board in 4 steps, and the card's showcase shows that text rendering correctly.

For newcomers and experts: Start in ComfyUI if you already use it, add the LoRA with the model-only loader and ask for a sign with a few words on it. Four steps with guidance at 1 is the whole recipe.

Version 2 is trained directly on the release schedule, so it works with the stock sampler instead of a custom one. The card says it was distilled with data-space DMD and started from an 8-step distilled model. On one A100, it reports about 11 seconds per 1024 by 1024 image. Applied to MOSS-Image-Base, the output is slightly softer.

Write a prompt
Load the one LoRA file
Sample 4 steps with the Euler sampler
A finished image comes back

Good to know:

• License: Apache 2.0, and the card says commercial use is fine
• Cost to run: none named, though you need a capable GPU
• Data it reaches: downloads files from Hugging Face, then runs where you load it
• Not for: older v1.x adapters, which keep their own sampler in a separate folder

On Hugging Face, 13 November 2026

• Task: text to image
• 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: 73,131
• Likes: 96
• License: apache-2.0
• Based on: OpenMOSS-Team/MOSS-Image-SFT
• Quantized versions: 1
• Tags: diffusers, safetensors, moss-image, text-to-image, turbo, few-step, lora, ml-intern

Links

• Model: https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps
• Files: https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download akhaliq/MOSS-Image-SFT-4-steps

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