Feedback
freeTrialImage.bannerPity
freeTrialImage.upgradeUnlock
- ✓freeTrialImage.benefitHd
- ✓freeTrialImage.benefitWatermark
- ✓freeTrialImage.benefitUnlimited
MOSS-Image
OpenMOSS built MOSS-Image on a 5B-parameter core that reads 4,000-token prompts and follows JSON blueprints for placing subjects, lettering and frames.
AI Tools
Explore AI tools
MOSS-Image: Precision Layout Meets Text-to-Image
This 5B-parameter image suite reads layout instructions directly from your prompt.
- Element-by-Element PlacementAssign a frame to every object and caption so the final composition mirrors your JSON structure rather than relying on chance.
- Compact 5B Core, Long ContextExtended prompts stay within the model's 4,000-token window, and the same weights can be tuned for specialized domains.
- Two Apache-2.0 CheckpointsBase and SFT releases share the same architecture; the SFT build produces cleaner whiteboard lettering and more organized text blocks.
Set Up MOSS-Image Step by Step
Go from a fresh environment to a layout-driven render in three stages.
Questions About MOSS-Image
Answers on layout prompts, model variants, hardware needs and licensing for MOSS-Image.
How much control does a JSON layout give me?
Your JSON can define the overall scene, individual objects and visible text, with coordinates and dimensions for each element's frame.
What separates the Base and SFT checkpoints?
Both use the same architecture and sampling interface. SFT has extra fine-tuning on curated examples, which the project shows as tighter text grouping and a cleaner whiteboard composition.
What prompt length does the model accept?
It handles prompts up to 4,000 tokens, which leaves room for a detailed multi-element layout in JSON.
Which system requirements should I plan for?
You need Python 3.10+ and an NVIDIA GPU with a compatible PyTorch build. When VRAM is tight, use Sequential mode or lower the image size; Compile DiT is disabled by default and not available in that mode.
Is there a ComfyUI integration for MOSS-Image?
Yes—install the native plugin through ComfyUI Manager, then open the sample workflow. Inclusion in the official ComfyUI repository is still underway.
How is MOSS-Image licensed?
The model weights use Apache-2.0; pipeline code and third-party parts keep their respective licenses. For academic work, cite the project repository.
Put MOSS-Image to Work
Run the Diffusers pipeline or ComfyUI plugin, test a JSON layout, or refine your concept with the built-in prompt helper before you generate.