ComfyUI MOSS Image

Install guide · checked Oct 21, 2026

How to install MOSS-Image in ComfyUI

MOSS-Image runs on ComfyUI's core nodes. You need ComfyUI v0.42.0 or newer, three model files and a workflow; this page covers the portable build, a git install and Comfy Desktop.

1. Check your ComfyUI version

The startup log prints a line like the one below, and Settings → About shows the same number. MOSS-Image needs 0.42.0 or newer: support was merged on Oct 16 and first released in v0.42.0 on Oct 18. The ComfyUI docs say 0.41.3 is enough, but that release does not include it.

ComfyUI version: 0.42.1

2. Update or install ComfyUI

Windows portable

New install: download ComfyUI_windows_portable_nvidia.7z (about 2 GB) from the v0.42.1 release, extract it with 7-Zip and run run_nvidia_gpu.bat. Existing install: run update_comfyui_stable.bat in the update folder to move to the latest release; update_comfyui.bat moves to the latest commit instead. Both reinstall the Python requirements when they change.

update\update_comfyui_stable.bat

Manual (git) install

In the ComfyUI folder run git pull (or git checkout v0.42.1), then pip install -r requirements.txt with the same Python. Don't skip the second command: v0.42.0 pins comfyui-frontend-package 1.55.18 and comfyui-workflow-templates 0.11.89, and the Template Library reads its templates from that package.

git pull
pip install -r requirements.txt

Comfy Desktop

Desktop ships its own ComfyUI core, and its latest stable release (v1.1.7, Oct 13) came out before ComfyUI v0.42.0. Check Settings → About: if the core is older than 0.42.0, MOSS-Image won't load, so use a portable or git install until Desktop catches up.

3. Download the model files

You need one diffusion model, the Qwen3.5-9B text encoder and the FLUX.2 VAE. The template set below is about 13 GB; other precisions are in the GPU picker.

FileSizeDownload
moss_image_sft_int8_convrot.safetensors5.41 GBHugging FaceModelScope
qwen3.5_9b_w4a8_gptq.safetensors7.34 GBHugging FaceModelScope
flux2-vae_small.safetensors0.25 GBHugging FaceModelScope

ModelScope has the same files under Comfy-Org/MOSS-Image, with identical names and byte sizes (checked Oct 21). It is usually faster from mainland China.

Other precisions: pick files for my GPU

4. Put each file in its folder

Paths are relative to the ComfyUI folder. Restart ComfyUI after copying so the loaders list the new files.

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── moss_image_sft_int8_convrot.safetensors
    ├── text_encoders/
    │   └── qwen3.5_9b_w4a8_gptq.safetensors
    └── vae/
        └── flux2-vae_small.safetensors

5. Load a workflow

Open the Template Library and search “MOSS-Image”, or copy one of the four workflows from our workflow page and press Ctrl+V on the ComfyUI canvas. Each workflow lists its model files with download links, so ComfyUI tells you if one is missing.

Get the workflows

6. Run the first image

Keep the template defaults for the first run: 25 steps, CFG 2.5, er_sde sampler, about 1 megapixel. Describe the finished image; the prompt upsampler expands short prompts.

If something fails

“Could not detect model type”, dtype errors and the other common failures are collected with their fixes on the home page.

See error fixes

Sources: ComfyUI v0.42.1 release · portable update.py · v0.42.0 requirements.txt · Comfy Desktop releases · ComfyUI v0.42.0 · Comfy-Org/MOSS-Image · ModelScope Comfy-Org/MOSS-Image · docs.comfy.org