moss_image_base_dmd2

moss_image_base_dmd2

Makes images from text in 8 steps instead of 50, as a distilled LoRA for the MOSS-Image-Base model. For ComfyUI and Diffusers users who want faster image generation.

Made by data-archetype 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_base_dmd2 is a LoRA with rank 128 for MOSS-Image-Base, made by distribution matching. It lets the base model draw an image in 8 passes with no negative prompt and a guidance scale of 1. The base model needs 50 guided steps for the same job.

Who it suits: It suits people who already generate images with MOSS-Image-Base and want results in fewer steps.

Industry | Who uses it | How they use it | Example
Design | Image creators | Draft pictures quickly with fewer steps | The card's goose by a rocky pond, made in 8 steps
Software | Diffusers developers | Load the adapter into a pipeline | A script that generates pictures from prompts
Research | Model testers | Compare distilled and base sampling | 8 steps against 50 guided steps

People who need a different sampler can skip it, since the card says the LoRA is trained for one sampler and others give worse results.

Where to use it:

• ComfyUI, through its native MOSS-Image support
• Diffusers, with the PEFT adapter the repository holds
• Windows, macOS or Linux, where you download the files with Python

Use case | Who | What they get back
Fast text to image | Image creators | A picture of about 1 megapixel from 8 steps
ComfyUI workflow | ComfyUI users | A ready workflow to drag in and run
Combining LoRAs | Advanced users | The LoRA loaded beside other LoRAs or MOSS-Image fine-tunes at strength 1

How to set it up: You need these first:

• Python and the huggingface_hub package
• ComfyUI with MOSS-Image support, or Diffusers plus a PyTorch install that fits your GPU driver
• The base MOSS-Image model, text encoder and VAE files from Comfy-Org, which the card names

1. Install the Hugging Face hub package and download the files.
2. For ComfyUI, put the LoRA in the loras folder and the base files in their own model folders.
3. For Diffusers, install the packages with the versions below.

pip install -U huggingface_hub
hf download data-archetype/moss_image_base_dmd2

pip install "diffusers>=0.38" "transformers>=5.3" "accelerate>=1.10" "safetensors>=0.6" "peft>=0.21" "Pillow>=10"

How to use it: In ComfyUI, drag the example image or the workflow file into the window. It sets 8 steps, a guidance scale of 1 and the right shift. Load the LoRA with the bypass loader, not the standard one.

Tool | What it does
LoraLoaderBypassModelOnly | Applies the LoRA without merging it into the bf16 weights
ModelSamplingAuraFlow | Sets the shift to the value the LoRA was trained with
lcm sampler | Takes a clean estimate at each step and adds fresh noise
simple scheduler | Spaces the 8 steps for the LoRA

Example: You load the workflow and write a prompt for a goose resting by a rocky pond. The card's own example of that scene was made in 8 steps at 1152 by 896 with seed 42. You get one picture back after 8 passes.

For newcomers and experts: A LoRA is a small add-on that changes how a big model behaves. This one teaches the base model to finish a picture in 8 steps. A first small thing to try is the card's example image, which carries its own workflow.

For experts, the sampler jumps to the model's clean estimate at each step and then re-noises it with fresh noise. The flow shift is 4.4817, while ComfyUI's default for MOSS-Image is 5. To change the shift, use ModelSamplingAuraFlow or ModelSamplingFlux. ModelSamplingSD3 handles the timesteps in another way, and the result is only noise. The standard LoRA loader folds the change into bf16 weights. The change is smaller than bf16 can round, so most of it disappears.

Load the base MOSS-Image model
Add the LoRA with the bypass loader
Set 8 steps, guidance 1 and the shift
Run the lcm sampler
Get the image

Good to know:

Title: Steps per image
This LoRA: 8 steps
Base model: 50 steps

• Size: rank 128 LoRA in Safetensors format, for images of about 1 megapixel
• Languages: English and Chinese are listed
• License: apache-2.0
• Cost to run: none named, no API key or account
• Data it reaches: it downloads files from Hugging Face and works on the prompts you give it
• Not for: other samplers, which give worse results, or the standard LoRA loader, which loses most of the change

On Hugging Face, 13 November 2026

• Task: text to image
• Library: peft
• Formats: Safetensors
• Downloads in the last 30 days: 0
• Likes: 23
• License: apache-2.0
• Languages: en, zh
• Based on: OpenMOSS-Team/MOSS-Image-Base
• Tags: peft, safetensors, moss-image, diffusers, lora, text-to-image, distillation

Links

• Model: https://huggingface.co/data-archetype/moss_image_base_dmd2
• Files: https://huggingface.co/data-archetype/moss_image_base_dmd2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download data-archetype/moss_image_base_dmd2

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