Download MOSS-Image and run it locally

MOSS-Image’s weights are free to download under Apache-2.0. Choose the route that matches how you work: ComfyUI for a visual workflow, Diffusers for Python, or the Gradio app for a quick local UI. Not ready to install anything? Generate online first.

Official downloads

Checkpoint	Hugging Face	ModelScope	Use it for
MOSS-Image-SFT	OpenMOSS-Team/MOSS-Image-SFT	openmoss/MOSS-Image-SFT	Everyday generation
MOSS-Image-Base	OpenMOSS-Team/MOSS-Image-Base	openmoss/MOSS-Image-Base	LoRA training, tag prompts
ComfyUI single files	Comfy-Org/MOSS-Image	—	ComfyUI (int8 and bf16)

The Hugging Face repositories are Diffusers directories: keep the full folder structure (transformer, text encoder, tokenizer, VAE, scheduler and the custom pipeline file). Users in mainland China can download faster from ModelScope.

Requirements

• GPU: an NVIDIA GPU with CUDA. OpenMOSS does not publish a minimum VRAM figure; the pipeline holds a 5B generator and a 9B text encoder, so plan memory accordingly or use the lower-memory options below.
• Software: Python 3.10+ and a PyTorch build that matches your CUDA version.
• Disk: the Diffusers repositories are tens of gigabytes; the ComfyUI int8 set is much smaller.

Install the official pipeline

git clone https://github.com/OpenMOSS/MOSS-Image.git
cd MOSS-Image
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Install a PyTorch build for your CUDA version before or after requirements.txt if pip picks the wrong one.

Generate with Diffusers

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "OpenMOSS-Team/MOSS-Image-SFT",
    custom_pipeline="pipeline_moss_image",
    trust_remote_code=True,
    torch_dtype={"default": torch.bfloat16, "vae": torch.float32},
)
pipe.scheduler = type(pipe.scheduler).from_config(pipe.scheduler.config)
pipe.to("cuda")

image = pipe(
    'A minimalist café poster on cream paper. The large title reads "SLOW COFFEE".',
    height=1024,
    width=1024,
    num_inference_steps=50,
    guidance_scale=3.0,
    uncond_skip_layers=[22],
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("poster.png")

• Width and height must be positive multiples of 16; 1–2 MP gives the best results.
• uncond_skip_layers=[22] is the official skip-layer guidance setting.
• The VAE runs in float32 for quality; everything else in bfloat16.

To use a layout, pass the JSON as a string: pipe(json.dumps(layout, ensure_ascii=False), height=1024, width=1024, ...). Keep the JSON coordinates consistent with the output size. See the layout guide.

Run the Gradio app

pip install -e '.[ui]'
python gradio_app.py --model OpenMOSS-Team/MOSS-Image-SFT

Paste a prompt or layout JSON into the Prompt field. Resolution and aspect-ratio controls calculate sizes aligned to 16 pixels for you.

Memory modes

The Gradio app’s Advanced settings expose three memory modes:

Mode	Behavior	When to use
GPU resident	Keeps the whole model on the GPU	Fastest; large GPUs
Component	Moves components between CPU and GPU as needed	Mid-size GPUs
Sequential · fallback	Lowest memory, slowest	When you run out of memory

If memory is still short, reduce the image size. Compile DiT is off by default and unavailable in Sequential mode. For the lightest local setup, use ComfyUI with the int8 diffusion model and the 4-bit Qwen3.5-9B encoder.

Faster and lighter community options

• FP16-fixed weights for GPUs where bf16 is slow or unsupported.
• 8-step DMD LoRA to cut sampling from 50 steps to 8.
• SGLang and vLLM-Omni had day-0 support for serving MOSS-Image in production.

Troubleshooting

• trust_remote_code error — the custom pipeline file must be present; download the full repository.
• Dimension error — width and height must be multiples of 16.
• CUDA out of memory — try Component or Sequential mode, 1 MP output, or ComfyUI int8.
• Washed-out or undercooked images — use 50 steps, CFG 3, shift 5 and skip layer 22.

Commands and settings checked against the official README and user guide on 26 October 2026.