MOSS Image Studio
Back to blog

MOSS-Image in ComfyUI: native workflow, files and settings (with sources)

Oct 22, 2026Studio notes

How MOSS-Image runs in ComfyUI: which ComfyUI version you need, which model files go where, the two official workflows and their settings. Every claim is tied to a primary source.

MOSS-Image in ComfyUI: native workflow, files and settings (with sources)

MOSS-Image can run inside ComfyUI in two different ways, and most guides blur them together. This post separates them, lists the exact files and settings, and links every claim to a primary source. Where something is not documented, we say so instead of guessing.

Short version

  • Since ComfyUI v0.42.0 (released 18 October 2026) MOSS-Image is supported natively. No plugin is needed for that path. (ComfyUI releases)
  • The official OpenMOSS repository also ships a separate plugin that loads a local Hugging Face / Diffusers model folder. (OpenMOSS/MOSS-Image, comfyui/README)
  • The native workflow template uses faster settings (25 steps, CFG 2.5) than the official defaults (50 steps, CFG 3). Both are legitimate; they trade speed for fidelity.

The model, from the primary source

According to the OpenMOSS MOSS-Image page and the official GitHub repository, MOSS-Image was released on 16 October 2026. It is a text-to-image diffusion model with a 5B-parameter backbone, comes as Base and SFT, and its weights are Apache-2.0 on Hugging Face. Supported ways to run it are Python with Diffusers, Gradio and ComfyUI.

A note on "editing": the OpenMOSS page says layouts support controlled editing. The repository README clarifies what that means: you edit the text, the element descriptions and the bounding boxes of a layout, then generate again. It is layout editing, not image-to-image editing. We found no statement of image input in the official material.

Path A: native ComfyUI (recommended)

Which ComfyUI

The MOSS-Image support was merged on 16 October 2026 in pull request 17095 and first shipped in release v0.42.0. The follow-up release v0.42.1 fixes crashes some users hit when running the MOSS-Image template workflow. The OpenMOSS README, written on 19 October, still warns that v0.41.3 does not contain the support. So update ComfyUI first. (ComfyUI releases)

Which files, and where

The Comfy-Org team publishes repackaged files in the Comfy-Org/MOSS-Image repository on Hugging Face (Apache-2.0). Place them like this:

  • ComfyUI/models/diffusion_models/: moss_image_sft_bf16, moss_image_sft_int8_convrot, moss_image_base_bf16 or moss_image_base_int8_convrot (.safetensors)
  • ComfyUI/models/text_encoders/: qwen3.5_9b_bf16, qwen3.5_9b_int8_convrot or qwen3.5_9b_w4a8_gptq
  • ComfyUI/models/vae/: flux2-vae or flux2-vae_small
  • ComfyUI/models/loras/: moss_image_color_depth_control (optional, a community colour/depth control LoRA listed in the same repository)

The ComfyUI-Wiki news post lists the sizes as about 10.2 GB for the bf16 diffusion models and about 5.4 GB for the int8 ones, with the text encoders around 7.3 to 9.8 GB. That is a third-party summary of the same repository, so check the file sizes on Hugging Face before you plan disk space.

What nobody publishes: a minimum VRAM figure. Neither the OpenMOSS README nor the ComfyUI template states one, and the third-party guides we read say the same. The smaller int8 / w4a8 files exist to reduce memory use, but how much you save on your card is something you will have to measure.

Which workflow

In ComfyUI open the template browser and pick the MOSS-Image text-to-image template. It comes from the Comfy-Org/workflow_templates repository (image_moss_image_t2i.json, with a Base variant image_moss_image_base_t2i.json). Its default settings, which we read from the template file itself:

  • Diffusion model moss_image_sft_int8_convrot, text encoder qwen3.5_9b_w4a8_gptq, VAE flux2-vae_small
  • 1024 × 1024, 25 steps, CFG 2.5, sampler er_sde, scheduler simple
  • Negative prompt: nsfw, worst quality, bad anatomy, watermark, jpeg artifact
  • Skip-layer guidance (SLG) at layer 22 is present but switched off in the template; the template notes say turning it off looks more realistic but can collapse details
  • Recommended total pixels: 1M, 1.5M or 2M

One more detail from the template notes that is easy to misread: the optional ref_image input goes to the language model for prompt enhancement only. The notes state that the image model cannot use it directly. It is not an image-to-image input.

The official defaults

If you want the settings OpenMOSS itself recommends, use the README values: flow-matching Euler, 50 steps, CFG 3 (3 to 6 works), shift 5, negative-branch skipping at layer 22. In the repository's own native workflow the sampler is Euler with the sgm_uniform scheduler at 1024 × 1024.

Path B: the OpenMOSS plugin

This path is for people who already have the Diffusers-format model folder. From the plugin README:

  • Install by copying comfyui/moss-image/ to ComfyUI/custom_nodes/moss-image/ and restarting, or run python comfyui/install.py --comfy-root /path/to/ComfyUI. A local/setup.sh and local/run.sh pair can also set up a separate ComfyUI checkout.
  • The loader node takes a local directory path, not a Hugging Face ID. The folder must contain transformer/, text_encoder/, tokenizer/ and vae/.
  • Nodes provided include the MOSS-Image HF Loader and a MOSS-Image Layout Prompt node for JSON layout prompting.
  • The example workflows are moss-image.json and an API version, moss-image-api.json.
  • Do not reuse a plugin workflow in the native path by only swapping the loader. The native model uses a different timestep multiplier, so replace the plugin's ModelSamplingSD3 node with ModelSamplingAuraFlow (multiplier 1).

How the pieces connect

Read the flow as one line, left to right:

  1. Prompt text (English recommended, Chinese supported), optionally a JSON layout
  2. Text encoder (Qwen3.5-9B variants)
  3. Diffusion model (MOSS-Image Base or SFT) with sampler, scheduler, shift and optional skip-layer guidance
  4. VAE decode
  5. Save image

Facts that circulate but we could not confirm

  • Minimum GPU or VRAM numbers. Several third-party pages state them. None of the primary sources do, and one popular page says openly that it has not run the workflow on a local GPU. We do not repeat the numbers.
  • "MOSS-Image does inpainting or image editing in ComfyUI." Not stated in the official material. See the note on layout editing above.
  • "8-step turbo." A community distilled 8-step model exists. It is a community release, not an official turbo model.

When you would rather not install anything

If you only want to test a prompt, a hosted route is quicker. See Use MOSS-Image online without installing ComfyUI, and for layout prompts see JSON layout prompting. Our hosted version uses the official API's own sampling defaults, so results will not match a local 50-step run exactly.

Sources

Checked on 22 October 2026. This is an independent site and is not affiliated with OpenMOSS or Comfy-Org; the projects move quickly, so defer to the sources above if anything differs.