Editorial status (October 21, 2026): Documentation-based walkthrough using the official ComfyUI release notes and model repositories. We have not independently benchmarked image quality, GPU memory consumption or inference speed.
What is MOSS-Image?
MOSS-Image is a 5-billion-parameter diffusion-transformer image-generation model introduced on October 16, 2026. The team provides Base and SFT variants, with an emphasis on readable text, prompt-driven layout and spatial placement of objects. The official project describes prompts up to 4,096 tokens. These are project specifications and claims, not results measured by The Signal.
ComfyUI 0.42.1: why update?
ComfyUI v0.42.0 introduced MOSS-Image support. The October 19 v0.42.1 patch specifically addresses crashes affecting some users of the MOSS image template workflow. Back up custom nodes and existing workflows before upgrading; follow our ComfyUI installation and troubleshooting guide.
Method A: official ComfyUI model files
- Update ComfyUI to v0.42.1 or a later compatible version.
- Open the Comfy-Org MOSS-Image repository and check its README and current filenames before downloading. The repository supplies Base/SFT diffusion checkpoints in BF16 and INT8 variants.
- Put one selected
moss_image_* .safetensorscheckpoint (without the space in the actual filename) inComfyUI/models/diffusion_models/; the matching Qwen3.5 text encoder goes inmodels/text_encoders/, andflux2-vae.safetensorsgoes inmodels/vae/. - Load the current MOSS-Image example template from ComfyUI’s workflow templates, select the downloaded checkpoint, text encoder and VAE, and run a small test image before increasing resolution.
File-size reality check: the official packaged BF16 diffusion checkpoint is about 10.2 GB; the INT8 diffusion alternative is about 5.41 GB. These are download sizes, not proven total VRAM requirements. The text encoder, VAE, activations and workflow settings also consume memory. See the current model file listing.
Method B: upstream MOSS custom nodes
The upstream ComfyUI plugin supports loading a local Hugging Face/Diffusers model directory. Copy the repository’s comfyui/moss-image/ directory to ComfyUI/custom_nodes/moss-image/, restart, and import its moss-image.json workflow. Set the HF Loader’s model_path to a local model directory containing transformer/, text_encoder/, tokenizer/ and vae/. Do not mix this directory-based loader with the single-file checkpoint setup above; they are alternative installation paths.
Layout control: a practical test
Start with a simple poster prompt: “Minimalist travel poster. Headline at top: EXPLORE THE DESERT. A single sandstone mountain centered. Small subtitle at bottom: TATAOUINE.” Generate an image, then compare text accuracy, element placement and consistency over several seeds. For precise bounding-box syntax, follow the official MOSS user guide rather than inventing a layout format. Record image dimensions, model variant, seed, sampling steps and output artifacts.
Troubleshooting
- Workflow crashes: verify you are on ComfyUI v0.42.1 or later and retry the maintained template.
- Missing model: check the exact filename and model folder; restart or refresh the model list.
- Out of memory: try a smaller resolution or a supported quantized variant; a smaller download does not guarantee a particular GPU will run it.
- Missing custom nodes: identify whether the workflow uses native ComfyUI support or the upstream MOSS plugin; do not install redundant nodes without a reason.
- Bad text placement: compare Base versus SFT and use the official layout examples; results vary with prompt and settings.
License and commercial deployment
The upstream repository and Comfy-Org package indicate Apache-2.0 licensing. Verify the exact checkpoint’s model card and all dependent assets before production use; this is not legal advice.
Related reading
See our comparison of AI image generators and Qwen-Image-2.1-Turbo setup. These guides describe documented setups; they do not claim original hardware benchmarks.
