Draw MOSS-Image layouts without doing coordinate arithmetic
MOSS-Image's experimental layout prompts give you a direct way to place text, characters and objects. The tricky part is the coordinate conversion: the model prompt uses absolute pixels measured from the center of the image, with the vertical coordinate first.
We built MOSS Layout Studio, a visual editor that handles that conversion. Draw and resize boxes, enter the exact words you want, and export the layout JSON. The editor runs in your browser and prepares prompts without calling an image model.
The wireframe is a composition tool. You can see where a headline, illustration and subtitle sit before spending time on generation. It is especially handy when text needs its own box or an intentional line break.
Two examples make the conversion concrete:
| Canvas | Draft box, top-left pixels [x1, y1, x2, y2] |
Model box, centered pixels [y1, x1, y2, x2] |
|---|---|---|
| 1536 × 1024 | [176, 224, 800, 864] |
[-288, -592, 352, 32] |
| 1024 × 1024 | [128, 96, 896, 224] |
[-416, -384, -288, 384] |
In the first row, the canvas center is (768, 512). Subtract 512 from the top and bottom edges, subtract 768 from the left and right edges, then put the vertical pair first. The second row is a title box near the top of a square poster. These examples follow the official layout reference and user guide.
Studio keeps two exports separate. The editable draft stores dimensions and conventional top-left boxes. The model prompt contains the centered boxes and omits dimensions, matching the upstream compiler. Keep the same output width and height in your generation settings. When importing a compiled prompt, choose its original canvas dimensions first.
Chinese copy, emoji, quotes and line breaks remain literal text. Known optional palettes, tags and composition metadata survive import and export. Overlapping items are allowed. Unknown fields and duplicate JSON keys produce an error so the editor cannot silently drop part of a composition.
We checked the compiler behavior against the pinned OpenMOSS Python reference on 494 accepted and rejected inputs, with zero mismatches. The cases cover non-square canvases, fractional edges, Unicode, optional metadata and invalid layouts. Twelve separate module test groups cover import/export, editing bounds and resize behavior. These checks establish representation compatibility; image quality and layout adherence still depend on the model and generation settings.
Try MOSS Layout Studio, or inspect the source and checks. A useful next experiment: take the same composition, change only the placement of one text box, and compare the generated images at fixed settings. Share the two layout drafts alongside the images so someone else can reproduce the comparison.
MOSS Layout Studio is an independent alphaBell project. The format and reference compiler come from OpenMOSS/MOSS-Image at revision 52b87fca23acedc399d062d593bec388aee9926a.
