MOSS-Image JSON layouts, explained
Prompts describe what is in a picture, but not reliably where. "A title at the top and a cup on the left" often comes back with the cup on the right and the title over it. MOSS-Image accepts a JSON layout instead: a caption for the whole scene plus a box for each element.
How to fix it
- The shape of a layout. A summary caption for style, light and mood, then a flat list of items. Each item has a label (text, object or character), a caption describing only what is inside it, a bbox, and for text items the exact words.
- The coordinates. The model reads boxes as [y_min, x_min, y_max, x_max] in pixels measured from the middle of the canvas ("abs_centralized", "yxyx"). On a 1024 x 1024 canvas both axes run from -512 to 512. Our editor lets you draw boxes normally and converts them for you.
- Let AI draft it. Describe the picture in plain words and press Generate JSON with AI. You get a full layout on your canvas, sized to the shape you picked, ready to adjust by dragging.
- Make text boxes big enough. Glyphs have to fit at the chosen resolution. A short title needs a wide box; three lines need height. Cramped boxes are the main cause of misspelled layout text.
- Change one thing, keep the rest. Keep the same seed (Remix), change one box ("make the couch green", "change the poster text"), and generate again. Most of the picture stays the same, which is the closest thing MOSS has to editing.
- Bring your own JSON. The JSON view shows exactly what the model gets. Paste a layout from the official prompt-enhancer skill, ComfyUI or another tool and press Apply.
Try it
The Layout tab is set up for exactly this, with examples to start from and an AI helper that writes the prompt for you.