MOSS-Image ComfyUI Workflow

Download the workflow

This file mirrors the example published in the OpenMOSS repository. It uses the optional MOSS-Image HF-directory custom node and expects a local Hugging Face model directory.

Mirrored October 23, 2026. Compare it with the upstream file before using it in a long-lived setup.

What is inside the workflow

The graph contains nine nodes: the MOSS-Image HF loader, positive and negative prompt encoders, an empty Flux latent canvas, skip-layer guidance, SD3 model sampling, a KSampler, VAE decoding, and image saving. The included bilingual sample prompt is useful for checking that the model and text encoder loaded correctly.

This is a text-prompt workflow. MOSS-Image's structured JSON layout syntax belongs inside the prompt; there is no separate JSON-layout node in this file.

Install and run it

1. Install and start a current ComfyUI build.
2. Clone the official MOSS-Image repository. Copy comfyui/moss-image/ into ComfyUI/custom_nodes/moss-image/, then install the dependencies listed by that plugin.
3. Download the MOSS-Image Base Hugging Face repository and keep its component folders together, including the transformer, text encoder, tokenizer, and VAE.
4. Import the downloaded JSON into ComfyUI. In MOSSImageHFLoader, set model_path to that local model directory.
5. Queue one square test image before changing nodes or adding extensions.

Start from the verified defaults

The upstream example is configured for a 1024 × 1024 image with 50 steps, CFG 3, shift 5, skip layer 22, the Euler sampler, and the sgm_uniform scheduler. Keep the seed fixed while validating the installation, then change one setting at a time.

These are workflow defaults, not universal performance recommendations. Generation time and memory use depend on your GPU, precision, offloading, ComfyUI version, and any custom nodes you install.

Troubleshooting

• Missing MOSSImageHFLoader: confirm the optional custom-node folder and its dependencies are installed, then restart ComfyUI.
• Model path error: point the loader at the full Hugging Face model directory, not a single .safetensors file.
• Out of memory: lower the canvas dimensions and use ComfyUI's supported memory-management options for your hardware.
• Unknown nodes after an update: use a workflow that matches the native or optional plugin installation path you chose.
• Text is imperfect: shorten the requested wording, make placement explicit, and proofread every generated image.

Use MOSS-Image without a local setup

If you want to test a prompt or structured layout before downloading model files, the online studio provides three starter image credits after email verification. No card or subscription is required.