MMOSS Image AI

MOSS-Image ComfyUI workflow

Two routes exist as of 18 October 2026: an official OpenMOSS plugin, and native support merged into ComfyUI master. We have not run either workflow ourselves. The steps below come from the official documentation and the ComfyUI pull request, not from our own tests.

Diagram of a ComfyUI-style graph: model, prompt, sampler, image
Diagram of a node graph. We have not run the official workflow, so this is not a screenshot of ComfyUI.
Anime group scene, an OpenMOSS published output. OpenMOSS published MOSS-Image output, assets/images/004
Published output (assets/images/004). Not a ComfyUI screenshot. Source: OpenMOSS assets/images/004. © OpenMOSS Team, Apache-2.0. We resized the long edge to at most 1400px and saved WebP. Not made on this site.

Status at a glance

RouteStatus (18 Oct 2026)Model filesSource
A. Official OpenMOSS pluginAvailable. Search “MOSS-Image” in ComfyUI Manager, or copy the folder. An example workflow ships with it.Diffusers folder from Hugging Face (transformer, text encoder, tokenizer, VAE)Plugin README
B. Native ComfyUI supportMerged into master on 16 Oct 2026 (PR #17095). The tagged release v0.41.0 predates that merge, so you need a current master or nightly build until the next release.Comfy-Org single files: diffusion model, Qwen3.5-9B text encoder, FLUX.2 VAEPR #17095
C. Community repacksint8, int4, GGUF, and Turbo files appeared within two days. None are official.VariousHugging Face search

Route A, from the plugin README

  1. Install the nodes in ComfyUI Manager, or copy comfyui/moss-image/ into ComfyUI/custom_nodes/moss-image/ and restart.
  2. Download MOSS-Image-Base or SFT as a folder, about 27.2 GB. The loader reads local files only.
  3. Drag moss-image.json from example_workflows into ComfyUI.
  4. Set model_path on the MOSS-Image HF Loader node to that folder.
  5. The example is preset to 50 steps, CFG 3, shift 5, negative-branch skip at layer 22, Euler, 1024×1024.

The README notes that weight_dtype controls DiT precision, the VAE runs in FP32, and text encoding is capped at 4,096 tokens. ComfyUI's scheduler differs slightly from the standalone pipeline, so identical seeds will not match official images exactly.

Route B: native ComfyUI

  • A model type for the 5B DiT, with LoRA key mapping.
  • moss_image as a text-encoder type (Qwen3.5-9B).
  • An experimental MOSS-Image Layout Prompt node that writes the same centre-origin JSON as our layout builder.
ComfyUI/models/
  diffusion_models/  moss_image_sft_bf16.safetensors (10.16 GB)
                     or moss_image_sft_int8_convrot.safetensors (5.41 GB)
  text_encoders/     qwen3.5_9b_int8_convrot.safetensors (9.75 GB)
                     or qwen3.5_9b_w4a8_gptq.safetensors (7.34 GB)
  vae/               flux2-vae.safetensors (0.34 GB)

Sizes are from the Hugging Face file listing. Because we have not run this graph, we do not invent a node-by-node recipe.

VRAM

OpenMOSS has not published a minimum VRAM figure. Full-precision files total about 27.2 GB. Treat any “runs on X GB” claim as unverified until OpenMOSS or your own test confirms it.

Which route

  • Documented path today: Route A.
  • ComfyUI nightly, smaller int8 files, layout node: Route B.
  • No suitable GPU: use Generate or the online options.

Sources

Facts checked 2026-10-18.