MOSS-Image in ComfyUI: Setup Guide and Workflow Overview

Independent guide. This page is written by Verdapix, which is not affiliated with the OpenMOSS team or Comfy Org. It summarizes the public ComfyUI documentation and Hugging Face model pages as of October 20, 2026. File names and version numbers change — always check the official docs before downloading.

Don't want to install anything?

You can generate MOSS-Image pictures in your browser with our online generator. New accounts get 3 free images. If you prefer to run the model on your own GPU, read on.

Two ways to run MOSS-Image in ComfyUI

1. Native ComfyUI templates (recommended)

The ComfyUI documentation includes a MOSS-Image tutorial with built-in workflow templates. According to the docs:

• You need ComfyUI 0.41.3 or later.
• In ComfyUI, open Templates, search for MOSS-Image, and pick a workflow such as MOSS-Image: Text to Image.
• The templates use the SFT checkpoint by default. There is also a Base variant and a layout-guided workflow where you draw bounding boxes.
• The text-to-image templates include a prompt upsampler that expands a short idea into a full caption before sampling.

2. The official MOSS-Image plugin

The OpenMOSS GitHub repository says you can install a native MOSS-Image plugin by searching for MOSS-Image in ComfyUI Manager. The plugin loads a Diffusers-format model folder.

Model files and where they go

The repackaged files for ComfyUI are hosted in the Comfy-Org MOSS-Image repository on Hugging Face. The documented layout is:

ComfyUI/
└── models/
    ├── diffusion_models/
    │   ├── moss_image_sft_int8_convrot.safetensors   (or the bf16 version)
    │   └── moss_image_base_int8_convrot.safetensors  (Base, optional)
    ├── text_encoders/
    │   └── qwen3.5_9b_int8_convrot.safetensors       (or bf16 / w4a8 versions)
    └── vae/
        └── flux2-vae_small.safetensors

Notes:

• The text encoder is a Qwen3.5 9B model that ComfyUI loads with type moss_image.
• The VAE is the FLUX.2 small VAE.
• int8 files need less VRAM than bf16 files. We have not confirmed exact VRAM requirements for each combination — check the model cards and community reports for your GPU.

Recommended settings

The official examples use about 50 sampling steps and a guidance scale of 3. Image width and height should be multiples of 16. The model accepts long prompts (up to about 4K tokens, according to the project). Community "turbo" LoRAs that cut the step count exist. We have not tested their quality.

Layout-guided generation

In the layout workflow you draw a box for each element and give it a caption. Boxes marked as text render that text inside the box. A layout-prompt node converts the boxes into the JSON layout format that MOSS-Image understands. This is the most precise way to place headlines and logos.

Common issues

The MOSS-Image templates don't appear

Update ComfyUI to 0.41.3 or later, restart it, then search the templates again.

Missing model or loader errors

Check that each file is in the correct folder (diffusion_models, text_encoders, vae) and that the text encoder is loaded with type moss_image.

Out-of-memory errors

Switch to the int8 diffusion model and text encoder, lower the resolution, or try our online generator instead.

FAQ

Is MOSS-Image officially supported in ComfyUI?

The ComfyUI documentation has a dedicated MOSS-Image tutorial with native templates. The OpenMOSS repository also offers its own plugin.

Do I need an API key to run it locally?

No. Local generation runs on your own GPU with the open weights. An API key is only needed for hosted services.

What is the difference between SFT and Base?

SFT is fine-tuned on curated data and, according to the model card, lays out text more cleanly. Base is the raw model, which is better suited for training LoRAs.

Can I use ComfyUI outputs commercially?

The weights are published under Apache-2.0. Check each file's model card, and make sure your outputs don't infringe other people's rights.