MOSS Image ComfyUI: setup, workflow and fixes

What is MOSS-Image? Parameters and hardware requirements

1. Choose native ComfyUI or the HF loader plugin

Begin by identifying which workflow you have. The official ComfyUI guide recommends native support in a sufficiently recent master/nightly build. It identifies commit e087d8f6 or later and notes that v0.41.3 did not contain that support as of October 19. Update the installation and its dependencies together, then restart before importing a native graph.

The optional repository plugin serves a different purpose: loading a complete local Hugging Face / Diffusers model directory. Keep that route for an existing plugin graph. To move to native loaders, import a native workflow rather than replacing one loader inside the old graph. In particular, do not retain the plugin’s ModelSamplingSD3 node unchanged; the official instructions describe a different native timestep multiplier.

These are two supported routes, not a requirement to install both. Record which one you chose alongside the graph. A workflow file and its matching loader implementation form a pair; the presence of a model filename alone does not establish compatibility.

Installation commands: choose one route

For an existing git checkout of ComfyUI, use its Python environment and update the code and dependencies together. Back up your workflows first.

cd /path/to/ComfyUI
git pull --ff-only
python -m pip install -r requirements.txt
python main.py

For the optional HF-directory plugin, run the official installer from a separately downloaded OpenMOSS repository and point it at your ComfyUI directory. This is an alternative route, not a native-workflow prerequisite.

git clone https://github.com/OpenMOSS/MOSS-Image.git
cd MOSS-Image
python comfyui/install.py --comfy-root /path/to/ComfyUI

2. Download MOSS-Image files into the matching folders

The Comfy-Org model repository provides repackaged diffusion weights, text encoders and VAEs. Choose the files requested by your workflow rather than downloading every precision variant. The native folder categories are:

ComfyUI/models/
  diffusion_models/  # MOSS-Image Base or SFT weights
  text_encoders/     # matching Qwen 3.5 text encoder
  vae/               # matching FLUX.2 VAE

Restart or refresh model discovery after placing the files. Check the full filename selected in each loader; a file inside an extra nested directory can be easy to overlook. The official HF listing is the source for current filenames and available precision variants. Do not rename unrelated weights to make a dropdown entry appear.

For the optional HF directory plugin, follow its separate instructions and select the complete local directory, including its component subdirectories. That loader does not use a Hub repository name as a substitute for downloaded local files. Use the official HF loader example workflow only with that route.

3. Generate one plain image before adding layout controls

In the native graph, verify the diffusion model, a CLIP Loader whose type is moss_image, and the VAE selection. For a first test, use a single image with a short prompt, no LoRAs, and no optional acceleration nodes. Save this baseline graph before trying a more complicated setup. The point of the first run is to establish that all components load and an image decodes.

A small orange ceramic vase on a pale stone table, one leafy branch, warm side light from a window, neutral cream wall, simple still-life photograph, no text.

For unaccelerated Base/SFT, the official SFT sampling table lists 50 steps, CFG 3, shift 5, flow-matching Euler and negative-branch skipping at layer 22. These are a reference baseline, not universal settings for community distilled adapters. Leave the imported workflow’s scheduler intact until the baseline works, and use the adapter author’s recommendations when adding one.

Save the seed, dimensions, selected checkpoint, precision and node versions. Change one setting at a time. Once a plain prompt runs, try a short quoted headline, then advance to the JSON layout workflow. This sequence makes it easier to distinguish a loading problem from a prompt or geometry problem.

4. Troubleshoot the failure you can actually observe

Missing nodes after importing

Identify whether the graph expects native loaders or the optional HF plugin. Check the version and the exact missing node names before adding more custom nodes.

Model missing from a dropdown

Compare its folder, extension and filename with the workflow’s loader category. Refresh model discovery, then select the intended file explicitly.

A graph loads but the image looks wrong

Return to the matching original graph. Check model, text encoder and VAE selections, then remove adapters. A mixed native/plugin sampling chain is a different issue from poor prompt wording.

Out of memory

Reduce the first test’s dimensions and batch size, close other GPU workloads, and review the installation’s supported offload or precision options. Do not assume a 5B backbone means the complete pipeline needs only 5B weights.

Words or layout are wrong

First verify the plain-image baseline. Then shorten the visible text, enlarge its box, and match layout coordinates to the exact output dimensions.

When asking for help, share a minimal graph, the error message, versions and component filenames. Remove private prompts and local personal paths first. We have not benchmarked your hardware, so this guide does not promise a minimum VRAM capacity or generation speed.

Common MOSS-Image questions

For free options, planned hosted packs and dated cloud billing details, see the public MOSS Image pricing page.

What to do after the first working image

Keep the successful graph as a clean reference. Duplicate it for experiments, give each change a descriptive name, and retain both good and bad outputs. If you change checkpoint, precision and prompt together, you cannot tell which change produced the result. Use the prompt guide to turn a broad idea into a testable brief, then compare several results against that same brief before settling on settings.