VarwoThe quickest way to get stuck with ComfyUI MOSS-Image is to mix a native workflow with a plugin loader. Choose one setup route, download its matching files, and check a small generation before changing advanced settings.
Get ComfyUI-format models and workflows · Read upstream integration instructions
Documentation checked October 23, 2026. This guide has not been validated by a local ComfyUI installation or a GPU generation run.
Choose native support or the HF-directory plugin
The MOSS-Image ComfyUI README recommends native support on a compatible ComfyUI revision. It also documents an optional plugin for loading a local Hugging Face/Diffusers directory. These are different routes, not interchangeable sets of nodes.
| You have… | Choose… | Keep together |
|---|---|---|
| A ComfyUI build containing native MOSS-Image support | Native workflow | ComfyUI-format components and native loaders |
| An existing plugin workflow or a local Diffusers model directory | HF-directory plugin | Plugin nodes, its example workflow, and the complete model directory |
| A workflow that reports missing nodes | Diagnose its route first | Do not install unrelated node packs to guess your way past the error |
Route A: native ComfyUI support
1. Check your ComfyUI revision
Upstream identifies commit e087d8f6 or later as containing the native integration and says release v0.41.3 did not contain it as of October 19. Check the build you actually run and its current dependencies instead of assuming every packaged release includes the feature.
Back up your workflows and environment before an update. If you depend on other custom nodes, evaluate compatibility in a separate environment rather than changing a working production setup blindly.
2. Download components into their model folders
The Comfy-Org model card provides filenames and linked native workflows. A documented BF16 Base combination uses:
ComfyUI/models/
diffusion_models/moss_image_base_bf16.safetensors
text_encoders/qwen3.5_9b_bf16.safetensors
vae/flux2-vae.safetensorsUse the files expected by your selected workflow. Alternative precisions and variants are listed upstream; do not infer that a smaller file is automatically compatible with every loader or device.
3. Import a matching workflow
Open a native workflow linked from the model card, import it into ComfyUI, and select the downloaded filenames in its loaders. Native loading uses Load Diffusion Model, CLIP Loader with moss_image type, and Load VAE.
Start with the workflow's own defaults. Confirm the queue completes and saves an output before adding LoRAs, layout control, or other nodes. There is no downloadable workflow generated by this page; the button above leads to upstream files.
Route B: load a local HF model directory with the plugin
If that is your intended route, obtain the upstream repository and review its optional plugin installer. From the repository root:
python comfyui/install.py --comfy-root /absolute/path/to/ComfyUIReplace the path with your existing installation. The installer links the node package and copies its example workflow when absent. Restart ComfyUI, then follow the plugin's upstream loading instructions with your complete local model directory.
This command is shown for the reader to review; it has not been run here. Model installation and GPU use happen in your environment.
Do not mix the two workflow types
Upstream warns that switching only the loader is insufficient: sampling conventions also differ. Use a complete native workflow when moving away from the plugin. In particular, do not carry the plugin's ModelSamplingSD3 configuration into a native graph unchanged. See the migration details.
A first-image check that catches real problems
Use one short prompt and record the exact workflow file, ComfyUI revision, model filenames, precision, dimensions, seed, and memory settings. For example:
A cream-colored square poster with a blue teapot in the center.
At the top, render exactly "TEA TIME" in large dark lettering.
Keep the background simple and add no other text.Check that an image is saved, the object appears, and the text is readable. A successful queue only proves execution; it does not prove the image met the request.
Troubleshooting checklist
| Problem | First diagnostic step |
|---|---|
| Unknown loader or missing node | Identify native vs plugin workflow; check installed revision |
| Model absent from the selector | Verify download completion, filename, folder, and model refresh/restart |
| HF Loader cannot find the model | Check the complete local directory, not just a Hub ID or one weight file |
| GPU out of memory | Record the failing stage and current dimensions; evaluate supported memory options one at a time |
| Queue succeeds but output is poor | Return to the matching upstream workflow and change only one parameter per comparison |
Before asking for help, collect a small reproducible workflow and the relevant error text. Do not share private prompts, credentials, or unrelated files in a diagnostic bundle.
Frequently asked questions
Do I always need a custom node?
Not for the documented native route. You need a compatible build and matching files. The HF-directory plugin serves a different loading workflow.
Where is the workflow download?
Use the workflow links on the Comfy-Org model card, or the plugin example linked from the upstream integration README. Choose the one that matches your route.
What GPU or VRAM is guaranteed to work?
This guide has no hardware benchmark or universal minimum-memory guarantee. Evaluate the exact model components, precision, resolution, and environment against current upstream requirements.
Is there an online generator on this page?
This is an installation guide. For a broader overview and a project-linked demo, see the MOSS-Image model guide.