A more reliable approach is to plan the layout before you generate, and to express that plan as bounding boxes: rectangles on the canvas that say "this exact text goes here, at roughly this size." Here is how to think about it, whether you hand-write the layout or use a visual tool.

Start from the content, not the picture

Before drawing a single box, write down every piece of text that must appear, exactly as it should be spelled. For an event poster that might be:

• Headline: "NIGHT MARKET"
• Date and time: "Fri 6–11 PM"
• Location: "Riverside Plaza"

Then rank them. The headline is level one, the date is level two, the location is level three. That hierarchy decides box size later. Skip this step and you get five equal boxes with no visual order.

Think of the canvas as coordinates

A bounding box is just four numbers. One common convention is [y_min, x_min, y_max, x_max]. Some layout formats measure from the top-left corner; others measure from the center of the canvas, so a 1024×1024 canvas runs from −512 to +512 on each axis.

A minimal layout item usually carries:

• a label (is this text, an object, or a character?)
• a short caption describing how it should look
• the exact text, if it is a text item
• the box itself

For example, a headline item might say: label text, caption "large dark-blue uppercase title," text "EXPLORE," and a box spanning the top fifth of the canvas, most of its width. A second item, labeled object, could hold the illustration in the middle.

Rules of thumb for sizing boxes

Big text needs big boxes. A box is a size instruction as much as a position instruction. Tiny boxes produce tiny letters, and a wide or long title squeezed into a narrow box gets compressed or cut. If a headline is long, give it more width or split it across two lines.

One block of copy per box. Don't put a whole menu in one rectangle. One line per item, or one box for names and a separate column for prices, keeps each piece readable and easier to check.

Leave breathing room. Boxes that touch tend to produce colliding text, and text at the very edge may get cropped when printed or posted.

Keep position out of the captions. If a caption says "title at the top" but you have dragged the box to the bottom, the two instructions fight each other. Let the box carry position and the caption carry style.

Control line breaks explicitly. Generators insert line breaks on their own, not always where you want them. If a break matters, put it inside the quoted text.

Mixing layout and prose

You can combine both: one summary sentence for mood and style ("a minimalist travel poster with a cream background"), plus boxes for everything whose position or wording matters.

Check the result like a proofreader

Treat the first image as a draft:

1. Compare every name, number and date against your list, character by character.
2. Check that text landed in its box. If something drifted, nudge or resize the box and generate again rather than rewriting the whole prompt.
3. View the image at its real display or print size. Small text that looks fine zoomed in can be too thin on a phone screen or a printed menu.
4. When you add text later, make room for it in the layout. When you remove text, remove its description too, or the old words may come back.

Tools that do the box math for you

Hand-writing coordinates gets tedious fast. Visual editors let you drag rectangles and copy the resulting layout. The free, open-source MOSS-Image Layout Editor (Apache-2.0) works this way in the browser. If you would rather skip the JSON entirely, MOSS-Image AI has a layout mode where you drag boxes onto the canvas for the headline, price or logo, and the saved layout can be reopened later to move a box and regenerate.

Whatever the tool, the habit is the same: decide the words, rank them, give each its own box, and proofread. Text placement stops being luck and becomes a plan.