How MOSS-Image prompts work
MOSS-Image is an open 5B-parameter text-to-image model from OpenMOSS that can follow either plain prose or a structured Layout. This tool writes either form for you.
Caption
A caption is cohesive prose describing the finished picture: subject and action, setting, framing, medium, lighting, materials, palette and mood. Any words that should appear in the image are quoted exactly. Use it for ordinary scenes, photographs and illustrations.
Layout
A Layout is a JSON prompt with a global summary and a flat list of items — objects, characters and text — each with its own description and a bounding box. Use it when text must land in a specific place, for posters, multi-panel images, or when object positions matter.
- Canvas size is outside the prompt. Width and height must be multiples of 16, and you must generate at exactly the size the Layout was drawn for.
- Boxes are converted for you. Drafts use pixel boxes from the top-left corner; the model expects them centered on the canvas in
[y1, x1, y2, x2]order. The compiler here does that conversion the same way as the official one, rounding outward to whole pixels. - Text is literal. Each text item carries the exact characters to render, including line breaks, so spelling and punctuation survive.
Where to paste it
In ComfyUI, paste the result into the positive prompt of your MOSS-Image workflow. For a Layout prompt, set the latent width and height to the size shown above the result.