Vai al contenuto
MOSS Image AI

How to run MOSS-Image in ComfyUI - setup, settings and memory

22 ott 2026

Questo articolo è disponibile in inglese.

MOSS-Image runs in ComfyUI from day one. This guide covers the two ways to load it, the settings OpenMOSS recommends, and what to do when your GPU runs out of memory. If you would rather skip the setup, the online studio runs the same model with no install.

Checked against the official repository on October 22, 2026. ComfyUI changes quickly; when this guide and the repository disagree, follow the repository.

Option 1: native support (recommended)

ComfyUI added MOSS-Image natively. No plugin is needed for this route.

  1. Update ComfyUI to a master or nightly build that includes MOSS-Image support, and install its updated Python dependencies.
  2. Download the ComfyUI-format weights from the Comfy-Org MOSS-Image repository on Hugging Face and place each file in the folder its model card names (diffusion model, text encoder and VAE go in separate folders).
  3. Load them with the built-in loaders and connect them to a sampler, as you would for other diffusion transformer models.
  4. Set the sampler to the values below and queue a prompt.

Option 2: the official plugin

The MOSS-Image repository also ships a ComfyUI plugin. Use it when you want to load a local Hugging Face / Diffusers model directory directly, or to open the workflows built for it.

  1. In ComfyUI Manager, search for MOSS-Image and install the plugin, or copy the comfyui folder from the repository into custom_nodes.
  2. Download MOSS-Image-SFT (or Base) from Hugging Face or ModelScope.
  3. Load the example workflow from the plugin's example_workflows folder and point its loader at your model directory.

Recommended settings

These are the defaults from the model card:

Setting Value
Sampler Euler (flow matching)
Steps 50
CFG 3 (raise toward 4 to 6 for short prompts)
Shift 5
Skip-layer guidance Negative branch, layer 22
Negative prompt Empty, or: nsfw, worst quality, bad anatomy, watermark, jpeg artifact
Size 1, 1.5 or 2 megapixels; width and height in multiples of 16

Use SFT for normal generation and Base as the starting point for training.

Memory: what to expect

OpenMOSS does not publish a minimum VRAM figure, and the real number depends on precision, image size and offloading. As a guide, the bfloat16 weights are about 10 GB for the diffusion transformer and about 18 GB for the Qwen3.5-9B text encoder. To fit a consumer GPU:

  • Offload components. Load the text encoder, encode the prompt, then free it before sampling. The official Gradio app calls this component mode; its sequential mode trades more speed for even lower memory.
  • Use lower-precision weights. Quantised or FP16 variants shrink memory use. The community project MOSS-Image-FP16-fixed provides FP16-compatible transformer weights that avoid overflow issues.
  • Generate smaller first. Test at 1 MP (1024 × 1024) before going to 2 MP.

Speed

Fifty steps, each running the transformer twice, take time even on a large GPU. Two community projects reduce it:

  • MOSS-Image-SFT-8-steps-dmd: an 8-step distilled variant with ComfyUI LoRA adapters.
  • ComfyUI-Spectrum-MOSS: nodes that forecast hidden states to skip some computation during sampling.

Both trade some fidelity for speed; read their model cards before relying on them.

Common problems

  • "Dimensions must be multiples of 16." Round width and height to the nearest multiple of 16.
  • Out-of-memory at the text encoder. Offload it or use a lower-precision encoder file.
  • Vertical colour banding. A known SFT artefact in some outputs; the community MOSS-Image-SFT-DeBand LoRAs target it.
  • Text comes out garbled. Put the exact words in quotes, keep them short, and write line breaks into the prompt. See the prompt guide.

Local or online?

Run it locally if you want to train LoRAs, chain it into larger ComfyUI workflows or generate in bulk on hardware you already own. Use the online studio when you want results in seconds without downloading about 28 GB of weights: it runs the same model on data-centre GPUs, and new accounts get free credits.

Sources

La tua prossima immagine è a un prompt di distanza

Inizia a creare
Aiuto