Feedback
freeTrialImage.bannerPity
freeTrialImage.upgradeUnlock
- ✓freeTrialImage.benefitHd
- ✓freeTrialImage.benefitWatermark
- ✓freeTrialImage.benefitUnlimited
MOSS-Image
Explore MOSS-Image: a 5B text-to-image model with 4K-token prompts and JSON layout control that pins every object, label and box in place.
AI Tools
Explore AI tools
A Text-to-Image Model Built Around Layout
A compact five-billion-parameter family where composition is written down, not guessed at.
- Layout by InstructionAssign every subject and line of rendered text its own box, so the frame follows your JSON instead of guesswork.
- 5B Weights, 4K-Token ContextLengthy prompts fit without trimming, and the same weights can be fine-tuned for your own domain or task.
- Base and SFT CheckpointsTwo Apache-2.0 releases share one architecture; the SFT build keeps whiteboard lettering noticeably cleaner.
Running MOSS-Image Step by Step
From a fresh environment to a fully controlled layout render in three moves.
MOSS-Image: Common Questions
Answers on layout prompts, checkpoints, hardware needs and licensing for MOSS-Image.
What can I actually control with a layout prompt?
Your JSON names the scene, its objects and any rendered text, then gives each element a box that fixes its position and size.
How is MOSS-Image-SFT different from the Base release?
Both share one architecture and sampling interface; SFT is tuned on curated data and, in the project's own example, groups text more cleanly and produces a neater whiteboard.
How much prompt text fits?
Roughly 4K tokens, which leaves plenty of room for a detailed multi-element layout JSON.
Which hardware and software are required?
Python 3.10+ plus an NVIDIA GPU with a matching PyTorch build. Under memory pressure, switch to Sequential mode or lower the image size; Compile DiT stays off by default and is unavailable there.
Does it work inside ComfyUI?
It does. Install the native plugin through ComfyUI Manager, then open the sample workflow; official inclusion in the ComfyUI repository is still underway.
What licensing applies to weights and code?
Model weights ship under Apache-2.0, while the pipeline code and third-party pieces keep their own terms. If you publish research, cite the project repository.
Ready to Compose with MOSS-Image?
Set up the Diffusers pipeline or the ComfyUI plugin, then test a JSON layout of your own — or shape the idea first with the MOSS-Image Prompt Enhancer.