OpenMOSS · 5B DiT
An open-source text-to-image model from OpenMOSS with a 5B diffusion backbone, built for in-image text and layout control. Explore the official demo and learn how to run it locally.
Hosted by OpenMOSS on Hugging Face Spaces. Your prompts are submitted directly to that service. Availability, sign-in, quotas and queues are controlled by Hugging Face and the Space owner. This demo does not use your site credits.
Base is the foundation checkpoint for experimentation and fine-tuning. SFT is fine-tuned on curated data; OpenMOSS shows clearer text grouping in this whiteboard comparison. These are official examples, not results generated on this site.
These original starter prompts are not the generation records for the examples above. Copy one and paste it into the demo. Layout control is experimental; use the matching 1024 × 1024 canvas for the JSON example.
A minimalist coffee shop poster on a warm cream background. At the top, large dark green letters read "MORNING COFFEE". In the center, a ceramic cup of coffee with soft steam. At the bottom, smaller text reads "OPEN DAILY". Clear typography, generous spacing, flat editorial illustration.
{
"summary": {
"meta": [
"poster"
],
"canvas": "single",
"caption": "A cream poster with a green title and a ceramic coffee cup."
},
"items": [
{
"label": "text",
"caption": "Large green title at the top.",
"text": "MORNING COFFEE",
"bbox": [
-440,
-440,
-280,
440
]
},
{
"label": "object",
"caption": "A ceramic coffee cup with steam, centered.",
"bbox": [
-160,
-240,
320,
240
]
}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
}Use a Python environment and a compatible PyTorch / CUDA setup for your GPU. The 5B figure describes the diffusion backbone, not the full runtime memory requirement. Check upstream requirements for your hardware; if memory is insufficient, reduce resolution or use the official offload options.
Linux / macOS / WSL shell (a compatible GPU runtime is still required)
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
pip install -e '.[ui]'
python gradio_app.py --model OpenMOSS-Team/MOSS-Image-BaseOn Windows PowerShell, activate with .\.venv\Scripts\Activate.ps1 instead of source. The launch command below runs the Base Gradio UI locally.
Official setup and memory modesThis guide uses the official custom plugin with a complete Hugging Face / Diffusers directory. Do not mix it with workflows for separately repackaged component weights.
Clone the official repository, copy comfyui/moss-image/ into ComfyUI/custom_nodes/moss-image/, then restart ComfyUI. Alternatively, search for MOSS-Image in ComfyUI Manager.
Download MOSS-Image-Base to a local folder. Keep transformer/, text_encoder/, tokenizer/ and vae/ together. The loader expects a local directory, not a Hub repository name.
Save moss-image.json from the link below, drag it into ComfyUI, and set model_path on MOSS-Image HF Loader to the local model directory.
Edit positive and negative prompts, keep the initial settings below, and click Run. Verify model_path and restart ComfyUI if nodes are missing; lower resolution or adjust device/offload settings for memory errors.
1024 × 1024 · 50 steps · CFG 3 · Shift 5 · Euler / sgm_uniform · negative-branch skip layer 22. These settings are for this Base workflow, not a universal SFT preset.
Sources checked: October 20, 2026. Follow upstream docs for later changes.