FLUX 3 Image bounding boxes: layout and edit prompts

Most image models let you describe a scene. FLUX 3 Image also lets you say where each thing goes. This guide explains the bounding box mechanism as Black Forest Labs documents it, then gives three example prompts you can adapt in the generator.

How a box is written

A bounding box is four numbers, [top, left, bottom, right], on a grid that runs from 0 to 1000 in both directions regardless of the image's real size or aspect ratio. [0, 0, 1000, 1000] is the whole frame; [0, 0, 500, 1000] is the top half; [700, 600, 1000, 1000] is the bottom-right corner.

Each box has a short id and a description of what goes inside it. BFL's own published layout for a festival poster, for instance, has a text box [10, 200, 170, 800] across the top for the title, a dome in the middle, swimmers below it, and a crowd along the bottom edge. You put the scene prompt first and append the element list at the end; the model places each element inside its box.

Two practical notes from the docs. First, boxes are optional: for a plain generation or a single edit you can skip them entirely, and the model will expand your instruction and, for many edits, find and box the element itself. Second, whatever it rendered from, including any boxes it added, comes back as the expanded prompt alongside the image, so you can see what it understood.

Example 1: a poster with the title where you want it

Scene prompt, then the boxes. Note the quoted title: quoted text renders as a literal.

A risograph poster for a seaside night market, two inks, coral and deep teal,
grainy paper texture.

Elements:
- title: [40, 100, 180, 900] — the words "NIGHT MARKET" in a tall condensed
  sans-serif, coral ink
- lanterns: [200, 0, 520, 1000] — a string of paper lanterns across the frame
- stalls: [520, 80, 880, 920] — a row of market stalls with striped awnings
- date: [900, 300, 970, 700] — the line "Fridays, June to September" in small
  teal capitals

The title gets the top band, the date gets a thin strip at the bottom, and the illustration fills the middle. If the lanterns drift into the title, shrink their box so it starts below 180.

Example 2: a product page hero at 16:9

Boxes shine when a layout has to match a template. Here the product is centered with room for copy on the left.

A product photograph of a matte black ceramic pour-over coffee dripper on a
light oak counter, soft morning window light, shallow depth of field, 16:9.

Elements:
- headline: [150, 60, 300, 460] — the words "Slow mornings, sharp coffee" in a
  light serif, charcoal
- subline: [320, 60, 400, 460] — one line of small grey text reading "Hand-glazed
  in small batches"
- product: [100, 520, 900, 940] — the coffee dripper, centered, with a glass
  carafe beneath it
- steam: [60, 600, 200, 860] — a faint wisp of steam rising from the dripper

Run it with aspect ratio 16:9 so the grid maps onto a wide frame. Because the copy boxes stop at 460 and the product starts at 520, the text never overlaps the dripper.

Example 3: editing one spot in a finished image

The same boxes target an edit. Upload the image as a reference, name the element, give its box, and say what stays.

In this image, change only the red bicycle leaning against the wall to a pale
mint green. Keep everything else exactly unchanged: the brick wall, the shop
sign, the pavement, the lighting.

Elements:
- bicycle: [420, 300, 820, 620] — the bicycle, now pale mint green with the
  same frame, saddle, and tires

For a move, give the element a source box and a target box and say where it goes; for a removal, describe what should fill the box instead (the wall, the pavement). The published BFL examples do exactly this: recolor a moth to leopard print, turn a purple ribbon to metallic silk, change a mesa to polished obsidian, each with the rest of the frame anchored.

Habits that help

• Say what stays. "Keep everything else exactly unchanged" is in nearly every BFL example for a reason.
• One element per box. A box with three things in it gives the model three ways to place them.
• Leave gaps between text boxes and image boxes. Overlap is where lettering gets lost.
• Check the expanded prompt. If the result surprises you, the prompt the model actually rendered from tells you what it understood.

Each generation costs the same whether or not it carries boxes; the price follows the resolution, as explained in how credits work. If you are new to the model, start with what FLUX 3 Image is, or go straight to the generator and the FAQ.

Source: Black Forest Labs, docs.bfl.ai, FLUX 3 Image overview and editing pages, and bfl.ai/models/flux-3-image, read October 2, 2026. The example prompts above are ours, not BFL's.