Unofficial guide · checked Oct 21, 2026

ComfyUI MOSS Image setup guide

Run OpenMOSS's 5B MOSS-Image in ComfyUI: the three files to download, where they go, what fits your GPU, and fixes for the first errors.

Pick files for my GPU Fix setup errors

5B DiT · Qwen3.5-9B text encoder
Apache-2.0
Released Oct 16, 2026
Native in ComfyUI v0.42.0+
English and Chinese text rendering

Set up MOSS-Image in ComfyUI

Native support, no custom nodes. Four steps.

1
Update ComfyUI to v0.42.0 or newer
Support was merged on Oct 16 (commit e087d8f6) and first shipped in v0.42.0 on Oct 18. The ComfyUI docs say 0.41.3 is enough, but 0.41.3 does not contain the change and fails with “Could not detect model type”.
2
Download three files
A diffusion model, the Qwen3.5-9B text encoder and the FLUX.2 VAE, all from Comfy-Org/MOSS-Image on Hugging Face. The template set is about 13 GB; the picker below lists the exact files.
3
Open an official template
Open the Template Library and search “MOSS-Image”. There are four: text to image and layout-guided text to image, each for SFT and Base.
4
Write the prompt and run
Describe the finished image, not your request. The prompt upsampler is on by default and expands short prompts with the same text encoder; set enable_prompt_upsampler to false to send your prompt unchanged.

Full install guide

Sources: ComfyUI PR #17095 · ComfyUI v0.42.0 · OpenMOSS ComfyUI README · docs.comfy.org

Which files fit your GPU

Pick the model and your GPU memory. You get the exact files, the folder each one goes in, and a download command for each.

Model

GPU memory

The template set: int8 model, w4a8 text encoder, small VAE. When VRAM runs short, ComfyUI moves parts to system RAM, so 6–8 GB cards also run, more slowly.

Diffusion model

moss_image_sft_int8_convrot.safetensors

Put in ComfyUI/models/diffusion_models/Download:Hugging Face ModelScope

Text encoder

qwen3.5_9b_w4a8_gptq.safetensors

Put in ComfyUI/models/text_encoders/Download:Hugging Face ModelScope

VAE

flux2-vae_small.safetensors

Put in ComfyUI/models/vae/Download:Hugging Face ModelScope

Total download13.01 GB

Download commands

hf download Comfy-Org/MOSS-Image diffusion_models/moss_image_sft_int8_convrot.safetensors --local-dir ComfyUI/models
hf download Comfy-Org/MOSS-Image text_encoders/qwen3.5_9b_w4a8_gptq.safetensors --local-dir ComfyUI/models
hf download Comfy-Org/MOSS-Image vae/flux2-vae_small.safetensors --local-dir ComfyUI/models

Run them in the folder that contains ComfyUI. They need the Hugging Face CLI: pip install -U huggingface_hub.

If huggingface.co is slow or blocked where you are, use the ModelScope links (same files) or point the HF_ENDPOINT environment variable at a mirror before running the commands.

Official VRAM requirements are not published. These picks follow the file sizes and ComfyUI's automatic offloading.

All Comfy-Org files

File	Size	ComfyUI folder
moss_image_sft_int8_convrot.safetensors ModelScopetemplate default	5.41 GB	ComfyUI/models/diffusion_models/
moss_image_sft_bf16.safetensors ModelScope	10.16 GB	ComfyUI/models/diffusion_models/
moss_image_base_int8_convrot.safetensors ModelScopetemplate default	5.41 GB	ComfyUI/models/diffusion_models/
moss_image_base_bf16.safetensors ModelScope	10.16 GB	ComfyUI/models/diffusion_models/
qwen3.5_9b_w4a8_gptq.safetensors ModelScopetemplate default	7.34 GB	ComfyUI/models/text_encoders/
qwen3.5_9b_int8_convrot.safetensors ModelScope	9.75 GB	ComfyUI/models/text_encoders/
qwen3.5_9b_bf16.safetensors ModelScope	19.31 GB	ComfyUI/models/text_encoders/
flux2-vae_small.safetensors ModelScopetemplate default	0.25 GB	ComfyUI/models/vae/
flux2-vae.safetensors ModelScope	0.34 GB	ComfyUI/models/vae/

Sources: Comfy-Org/MOSS-Image · docs.comfy.org · Comfy-Org discussion #4

Official workflows and settings

All four templates run the same subgraph. Copy them into ComfyUI from our workflow page, or open them from the Template Library.

Copy or download the workflows

Recommended settings

Setting	ComfyUI template	OpenMOSS reference
Steps	25	50 or more
CFG	2.5	3 (3–6; higher for short prompts)
Sampler	er_sde	Euler (flow matching)
Shift	5 (model default)	5
Skip-layer guidance	On, layer 22	Layer 22
Resolution	1 MP by default	1–2 MP, sides in multiples of 16

To change shift, use ModelSamplingAuraFlow. ModelSamplingSD3 causes dtype errors with the int8 model.
Turning skip-layer guidance off can look more realistic but may collapse detail.

Sources: docs.comfy.org · OpenMOSS ComfyUI README · OpenMOSS user guide

Layout JSON, explained

MOSS-Image can place text and objects inside boxes you define. In ComfyUI, draw boxes with Create Bounding Boxes and the MOSS-Image Layout Prompt node turns them into this JSON.

bbox is [y_min, x_min, y_max, x_max] in pixels, measured from the center of the canvas (bbox_coord_space “abs_centralized”). On a 1024×1024 canvas each axis runs from −512 to 512; x grows to the right and y grows down. To convert ordinary top-left coordinates, subtract half the width from x and half the height from y.

Each item has a label (text, character or object), a caption and a bbox; text items also carry the text to render. Box coordinates must match the output size.

Official example (1024×1024)

{
  "summary": {
    "meta": ["poster"],
    "canvas": "single",
    "caption": "A minimalist travel poster with a cream background, a blue mountain illustration, and the title EXPLORE at the top."
  },
  "items": [
    {
      "label": "text",
      "caption": "Large dark-blue uppercase title centered at the top.",
      "text": "EXPLORE",
      "bbox": [-416, -384, -288, 384]
    },
    {
      "label": "object",
      "caption": "A flat blue mountain illustration in the lower center.",
      "bbox": [-192, -288, 384, 288]
    }
  ],
  "bbox_coord_space": "abs_centralized",
  "bbox_format": "yxyx"
}

Sources: OpenMOSS/MOSS-Image README · OpenMOSS user guide · docs.comfy.org

Errors and fixes

Collected from the Comfy-Org and OpenMOSS discussion threads and issues.

“Could not detect model type”, or moss_image is an invalid CLIP type

ComfyUI is older than v0.42.0; 0.41.3 has no MOSS-Image support. Update to v0.42.0 or later, or to nightly. Don't switch the CLIP type to Flux 2.

The layout template downloads the wrong model

In the official SFT layout template, the download link for moss_image_sft_int8_convrot.safetensors points to the Base file. If you downloaded through it, get moss_image_sft_int8_convrot.safetensors directly instead. Our copy of the workflow has the corrected link.

KSampler dtype error with the int8 model

There is a ModelSamplingSD3 node in the graph. Replace it with ModelSamplingAuraFlow or ModelSamplingFlux.

Which custom nodes do I need?

None. The Comfy-Org files run on core nodes. The OpenMOSS plugin is only for loading a Hugging Face/Diffusers folder.

AttributeError: 'NoneType' object has no attribute 'dtype'

The OpenMOSS plugin's loader fails on the Comfy-Org single files. Load them with the core Load Diffusion Model node instead.

Is there a GGUF version?

Not from Comfy-Org, which recommends int8 with automatic offloading instead. Community GGUFs exist, such as realrebelai/MOSS-Image_GGUFs.

Vertical color bands in SFT images

The community DeBand LoRA (MATLOWAI/MOSS-Image-SFT-DeBand), listed in the official README, reduces them.

TorchCompileModel breaks the workflow

It conflicts with the skip-layer guidance patches. Run without torch.compile.

Same seed, different image from the Python pipeline

Expected. ComfyUI snaps the scheduler differently, so seeds don't reproduce the standalone pipeline exactly.

Sources: Comfy-Org discussion #6 · Comfy-Org discussion #4 · OpenMOSS issue #10 · OpenMOSS ComfyUI README · OpenMOSS/MOSS-Image README

FAQ

Short answers, each based on the sources linked on this page.

Ready to run it?

Update ComfyUI to v0.42.0 or newer, download the three files, and open a MOSS-Image template.

Pick my files

ComfyUI MOSS Image setup guide

Set up MOSS-Image in ComfyUI

Update ComfyUI to v0.42.0 or newer

Download three files

Open an official template

Write the prompt and run