MOSS-Image-IDM-2-Steps

MOSS-Image-IDM-2-Steps

Makes images from text in two steps with no guidance pass, as a small adapter for MOSS-Image-SFT, plus a second adapter for light image edits. Made for people who run image generation locally.

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: This is a single LoRA adapter that lets the MOSS-Image-SFT image model finish a picture in 2 steps, with no classifier-free guidance. It is a community build trained with the loss mix from Intrinsic Distribution Matching. The repository also holds a second adapter for small edits.

Who it suits: It suits people who run text-to-image models on their own machine and want faster generation, including readable text inside pictures.

Industry | Who uses it | How they use it | Example
Design | Poster and sign designers | Draft text-heavy images quickly | Make a chalkboard menu titled MORNING BREW with three priced lines
Games | Game UI artists | Sketch interface looks from a prompt | Try a game UI concept at two steps
Marketing | Content creators | Produce quick comic, poster or anime-style visuals | Generate a neon bar sign with short lines of text
Research | Diffusion researchers | Study few-step distillation of a large image model | Compare two-step output with the 50-step teacher

If you need heavy edits such as turning a sketch into a photo, or colourizing, you can skip it.

Where to use it: It runs in a Python environment on your own computer, with the base MOSS-Image-SFT model.

• Python on Windows, macOS or Linux
• The diffusers library, below version 1.0, for the editing script
• The transformers library, up to the 5.20 series, for the same script

Use case | Who | What they get back
Text-to-image in two steps | Image makers | A picture with legible text, from the v1 checkpoint
Attribute edits | Editors | Changes such as hairstyle or expression, using the v2 adapter
Few-step research | Researchers | A reference adapter and evaluation grids to compare against

How to set it up: You need Python, the base MOSS-Image-SFT model, and a checkpoint from this repository. Pick v1 for text-to-image and v2 for edits.

1. Install the Hugging Face hub tools.
2. Download the files in this repository.
3. For text-to-image, use the v1 IDM export 750 checkpoint.
4. For edits, use the folded v2 adapter and its inference_edit.py script.

pip install -U huggingface_hub
hf download akhaliq/MOSS-Image-IDM-2-Steps

How to use it: Choose the checkpoint that matches the job. The card says v2 visibly weakened general text-to-image, so use v1 for that. Edits run locally through the inference_edit.py script that ships in the repository.

Checkpoint | Use for | Where
v1 IDM export 750 | Text-to-image, recommended | runs/idm/export/000750/
v2 folded adapter | Attribute edits | The repository root or adapter_v2

Tool | What it does
hf download | Fetches the repository files to your computer
inference_edit.py | Runs 2-step edits with the folded v2 adapter
v1 IDM export 750 | The recommended text-to-image adapter

Example: For text-to-image, you describe a chalkboard menu reading MORNING BREW with three priced lines, and the 2-step v1 adapter returns that picture. For an edit, you give a portrait and ask for "long wavy auburn hair", and the v2 adapter returns the portrait with the same person and the new hair.

For newcomers and experts: A LoRA adapter is a small add-on file that changes how a large model behaves. This one makes the image model finish in 2 steps instead of 50. A first try is to download the repository and look at the card's sample prompts, then write one like the chalkboard menu.

The lineage runs from the 50-step base through an 8-step and a 4-step adapter to this IDM 2-step one. Each image costs 2 forward passes of the diffusion transformer, against 100 for the base. The trade-off: the edit continuation hurt plain generation, and structural edits fail at 2 steps.

Start from the MOSS-Image-SFT base model
Add the 2-step LoRA adapter
Write your prompt with no guidance pass
Get the image after two steps

Good to know: The card compares the cost of one 1024 by 1024 image between the base and this adapter.

Title: DiT forwards per 1024 by 1024 image
MOSS-Image-SFT, 50 steps with CFG 3: 100 forwards
This adapter, 2 steps, no CFG: 2 forwards

• Editing limit: attribute edits such as expression or colour work, but sketch to photo, colorization and big style transfers do not.
• A 12-type held-out edit test scored about 1 in 12 on instruction following at 2 steps.
• For production editing, the card points to a 50-step edit setup instead.
• License: Apache 2.0.
• Cost to run: none named. It runs on your own machine.
• Data it reaches: the prompts and pictures you provide, handled locally.
• Not for: structural image transformations at 2 steps, or general text-to-image with the v2 adapter.

On Hugging Face, 14 November 2026

• Task: text to image
• Downloads in the last 30 days: 0
• Likes: 43
• License: apache-2.0
• Languages: en, zh
• Based on: OpenMOSS-Team/MOSS-Image-SFT
• Tags: moss-image, lora, text-rendering, text-to-image, distillation, few-step, ml-intern

Links

• Model: https://huggingface.co/akhaliq/MOSS-Image-IDM-2-Steps
• Files: https://huggingface.co/akhaliq/MOSS-Image-IDM-2-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-IDM-2-Steps

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