MOSS-Image has a native ComfyUI plugin maintained in the OpenMOSS repository. This tutorial follows the official installation instructions. We checked the documentation and workflow links; we have not run this workflow on a local GPU. The image shown on our homepage was generated in the official Space.
What you need
Use a working ComfyUI installation, a compatible Python/PyTorch/CUDA environment, and enough system RAM, disk and GPU memory for the chosen weights and memory settings. A 5B backbone is not a VRAM requirement for the whole pipeline. We do not yet have a verified minimum GPU specification to recommend.
If you only want to try a layout, use the browser editor and on-page generator first. The generator calls the official Space API with your authorized HF free account and uses its queue and quota.
Install the native plugin
The official project says you can search for MOSS-Image in ComfyUI Manager. Its repository also provides a manual path. Clone it, then run the installer with your existing ComfyUI path:
git clone https://github.com/OpenMOSS/MOSS-Image.git
cd MOSS-Image
python comfyui/install.py --comfy-root /path/to/ComfyUI
Replace /path/to/ComfyUI with your actual installation directory. Alternatively, copy the repository’s comfyui/moss-image/ directory to ComfyUI/custom_nodes/moss-image/. Restart ComfyUI after installation. The official README describes integration into the main ComfyUI repository as work in progress; use the plugin instructions for the current route.
Download the complete model directory
Start with MOSS-Image-Base. Keep the complete directory structure, including:
MOSS-Image-Base/
transformer/
text_encoder/
tokenizer/
vae/
The MOSS-Image HF Loader node expects model_path to point to that local directory. It does not download weights from a Hugging Face model ID. Do not enter only OpenMOSS-Team/MOSS-Image-Base as the local path.
Import the official workflow
Download the official moss-image.json workflow using GitHub’s raw download option, then drag the file into ComfyUI. Set the loader’s model_path, edit the positive and negative prompts, and run the workflow.
Its documented starting settings are 1024 × 1024, 50 steps, CFG 3, Shift 5, Euler with sgm_uniform, and negative-branch skip layer 22. These are model sampling recommendations, not a speed or quality guarantee. The plugin uses ComfyUI’s memory management; if memory is insufficient, follow the current official guidance and reduce output size before increasing workload.
Use a layout from this site
Build your composition in the layout editor, then copy its JSON into the workflow’s positive prompt. Match the workflow width and height to the exported canvas. For the homepage example, use 1024 × 1024 and seed 42. Download that exact input.
A layout JSON prompt and a ComfyUI workflow JSON are different files. Load the official workflow into ComfyUI, then paste the layout JSON into its prompt field. Dragging a layout prompt into the workflow canvas will not create the necessary nodes.
Troubleshooting
- Missing MOSS nodes: confirm the plugin is under the active installation’s
custom_nodes, restart ComfyUI, and read its startup errors. - Model not found: check the local path and the four model subdirectories; the loader does not fetch a Hub ID automatically.
- Incorrect layout positions: verify matching output dimensions and centered
yxyxcoordinates. See the coordinate guide. - Different result from the Space: the official plugin documentation warns that native ComfyUI sampling and the standalone pipeline differ. Matching the seed and parameters does not guarantee matching pixels.
- Misspelled or extra text: inspect every result. Layout control is experimental; our actual example contains an extra English line.
Official references
ComfyUI plugin README, model user guide, model project. Instructions checked October 17, 2026; follow upstream changes when installing later. This is an independent tutorial.