What is MOSS-Image, and who should try it?
OpenMOSS describes a 5B text-to-image family that accepts ordinary descriptions and structured layouts. Base and SFT share a sampling interface; SFT is fine-tuned on curated data. The project lists Apache-2.0 for weights, with separate terms for pipeline code and third-party components. Consider it when a poster, diagram or visual brief depends on where text and objects sit. This is a source-based guide, not a hands-on quality ranking.
Base vs SFT: compare your actual design brief
Start with one brief and keep canvas, seed and sampling settings recorded. Do not assume fine-tuning makes one checkpoint universally better. Our suggested comparison separates text accuracy, placement and visual quality so one impressive image cannot hide a repeated failure.
| Choice | What the project describes | What you should check |
|---|---|---|
| Base | Foundation text-to-image checkpoint | Whether it follows your subject, text and layout |
| SFT | Curated-data fine-tuning of Base | Whether its changes help your intended design |
| Either checkpoint | Same sampling interface | Exact artifact, settings, license and output review |
ComfyUI setup: choose native support or the plugin
The repository recommends native ComfyUI support containing commit e087d8f6 or later. Its documentation says v0.41.3 lacked this support as of October 19; check your installed revision rather than assuming every stable release includes it. Native files go in models/diffusion_models, models/text_encoders and models/vae. Use Load Diffusion Model, CLIP Loader with moss_image type, and Load VAE. Import a native example workflow when migrating. The older repository plugin loads a local Hugging Face/Diffusers directory and uses different nodes. In particular, carrying ModelSamplingSD3 unchanged from the plugin graph into a native graph is incorrect; the documentation specifies ModelSamplingAuraFlow for adjusting native shift. Save your working graph before changing environments.
- Record ComfyUI revision and dependency versions.
- Choose one documented workflow path.
- Match each downloaded component to that path.
- Load the supplied example before adding custom nodes or LoRAs.
- Keep a copy of the graph and exact model files used.
JSON layout coordinates: a worked conversion
The user guide calls layout control experimental. Its boxes use [y_min, x_min, y_max, x_max] and absolute pixel coordinates centered on the canvas. For our original 1024 × 1024 example, a conventional top-left box with x=100 to 900 and y=100 to 220 becomes [-412, -412, -292, 388]: subtract 512 from each coordinate, then put y first. A valid box only specifies intent; it does not prove correct output. Check both syntax and visual placement before increasing complexity.
| Field | Meaning | Common mistake |
|---|---|---|
| bbox_format: yxyx | Vertical bounds precede horizontal bounds | Using xyxy order |
| bbox_coord_space: abs_centralized | Pixel origin at canvas center | Using normalized 0–1 values |
| items[].text | Literal text to render | Describing the title without supplying its text |
| summary.caption | Whole-scene description | Contradicting the individual items |
Download an editable layout prompt
Use our original MOSS-Image layout prompt for a fictional FIELD NOTES poster at 1024 × 1024. This is a prompt JSON file, not a ComfyUI node graph or a tested model output. Load a compatible example workflow first, then use the JSON as layout input through the documented interface. Keep the canvas dimensions unchanged while checking coordinates. Replace only the title and scene description for your first variation; preserve a copy of the starting file so you can compare intent with output.
| If you see this | Check this first |
|---|---|
| moss_image is missing from CLIP Loader | Installed ComfyUI revision and dependencies; use a native-supporting build |
| Native graph fails after migrating from the plugin | Import a native example graph; do not swap only the loader |
| Model files do not appear | Component folder and loader type; refresh or restart the app |
| Layout appears misplaced | Centered pixel coordinates, yxyx order and the requested canvas |
| Text is misspelled | Exact literal text and line breaks; inspect output rather than trusting valid JSON |
First-run settings and memory expectations
The user guide recommends Euler, shift 5 and a CFG default of 3, with at least 50 steps for its preferred results. It requires manually entered dimensions to be positive multiples of 16. These are publisher recommendations, not our measured optimum. The project quick start uses Python 3.10+ and an NVIDIA GPU. We have no independently verified minimum VRAM figure: the 5B backbone excludes other pipeline components and working allocations. Keep the documented graph intact for your first run, then change one variable at a time.
A useful output test before adopting it
Create a fictional event poster with a short title, one object and one date. Before generating, write down the expected spelling, box placement and essential visual details. Compare Base and SFT on that same brief, retaining rejected outputs. Test plain language separately from JSON so you can tell whether layout control helped. This protocol is editorial advice, not an executed benchmark.
- Check title and date character by character.
- Inspect whether objects remain inside intended regions.
- Record model revision, seed, graph and canvas dimensions.
- Separate generation time from loading time.
- Count accepted outputs and manual correction effort.
Continue with safer downloads and detail preservation
Use the model download safety checklist before installing unfamiliar code. Read our upscaling versus img2img guide before changing an accepted image. For hardware planning, distinguish checkpoint size from full workflow memory in the local AI workstation guide.
Frequently asked questions
What is MOSS-Image?
A 5B text-to-image model family from OpenMOSS, with Base and SFT checkpoints and support for natural-language prompts and experimental JSON layouts.
Does MOSS-Image work in ComfyUI?
The official repository documents native ComfyUI support and a separate local-directory plugin. Use a compatible revision and the matching example workflow; their sampling nodes are not interchangeable.
How are MOSS-Image bounding boxes written?
The documented layout uses centered absolute pixel coordinates in yxyx order: y minimum, x minimum, y maximum, x maximum.
How much VRAM does MOSS-Image need?
We have not independently established a minimum. A 5B backbone is not the memory footprint of the full pipeline; test the exact components, precision, canvas and runtime.
Is the article thumbnail a MOSS-Image sample?
No. It is an original AI-generated editorial illustration, not a model output or test result.
Sources & further reading
Primary sources checked Oct 21, 2026. Vendor statements are attributed; editorial advice is our own.
- 1
- 2MOSS-Image official repository ↗OpenMOSS
- 3
- 4
Help us keep this useful. Report a factual correction →
