A good AI campaign image needs more than attractive lighting. It needs room for a headline, a product that sits in the right place, and a composition that survives the final crop. FLUX 3 Image bounding boxes offer a way to describe that layout explicitly instead of repeatedly asking a model to move something “a little to the right.”
Layout control is a timely topic for November campaign production. Teamday’s October 29 model roundup highlights FLUX 3 Image as a model to test for bounding-box layouts. Black Forest Labs’ documentation and independent coverage from The Frontier confirm the underlying feature: elements can be positioned using a normalized coordinate grid. The model launched on October 1; this is a practical guide to an existing capability, not a claim that it launched this week.
Below is a layout-first workflow for a seasonal campaign visual, including a sample prompt, a coordinate example, and an approval checklist. The examples are proposed starting points, not results from a hands-on benchmark.
What are bounding boxes in AI image generation?
A bounding box is a rectangle describing where an element should appear. Rather than relying only on “put the mug on the right,” you give the mug a named region and describe what belongs there.
In FLUX 3 Image, boxes use [top, left, bottom, right] coordinates, with each value running from 0 to 1000. The top-left corner is the origin. A box of [0, 0, 500, 500] covers the top-left quarter of the frame.
The order matters: this is not the common left-top-right-bottom convention. Swapping the first two coordinates can put an object in the wrong part of the image.
Black Forest Labs specifies that generation boxes belong in a JSON array appended to the text prompt, not in a separate box parameter. Each row has an id, a bbox, and a desc. The prose names the same IDs inside angle brackets.
Start with a campaign layout, not a longer prompt
Imagine a fictional winter campaign for an unbranded ceramic mug. The brief is a wide hero visual with the mug on the right and quiet space on the left for an offer added later in a design tool.
- Choose the final format. Begin with a 16:9 landscape composition. Do not plan a square layout and assume that cropping it will produce a good banner.
- Define the hierarchy. The mug is the visual focus. A small folded cloth supports it. The background stays simple enough for a readable headline.
- Reserve text space. Keep the left portion clear. Add prices, dates, logos, and legal copy as editable text after generation.
- Write acceptance criteria. One mug, one handle, a clean silhouette, a believable contact shadow, and no decorative objects entering the copy area.
This is different from marking up an existing photo for an edit: you are planning a new composition before generating it. If the image must depict an actual product for sale, use authorized product references in a supported editing workflow and verify its shape and details against the real item. A plausible invented mug is not accurate product photography.
A sample FLUX 3 Image layout prompt
Use the following as the prompt text and select a matching 16:9 aspect ratio in the supporting interface or request. The JSON is part of the prompt. This is a generation example, not the full API request body.
Create a restrained winter campaign photograph. An unbranded matte ivory ceramic mug <mug_1> sits on a warm stone surface against a softly lit beige background <background_1>. A folded dark green linen cloth <cloth_1> supports the mug. Soft daylight enters from the upper left. Keep the left 45 percent visually quiet for a headline added later. No generated text, logos, extra mugs, or seasonal decorations.
[
{
"id": "background_1",
"bbox": [0, 0, 1000, 1000],
"desc": "Warm beige studio background and stone tabletop, soft tonal transitions, uncluttered left side."
},
{
"id": "cloth_1",
"bbox": [700, 520, 920, 950],
"desc": "A neatly folded dark green linen cloth resting flat on the tabletop."
},
{
"id": "mug_1",
"bbox": [220, 590, 790, 880],
"desc": "One matte ivory ceramic mug with a single handle on its right, realistic rim, no writing or logo, natural contact shadow."
}
]The overlapping cloth and mug regions are intentional: the objects should form one scene. The background spans the canvas because it supports the entire composition. A box is a placement guide, however, not a clipping mask. BFL’s documentation warns that elements can extend beyond their boxes, so the left-side copy space still needs visual inspection.
Convert a design rectangle into coordinates
If you already have a wireframe, convert its pixel measurements into the 0–1000 grid:
- Top = top pixel position ÷ canvas height × 1000.
- Left = left pixel position ÷ canvas width × 1000.
- Bottom = bottom pixel position ÷ canvas height × 1000.
- Right = right pixel position ÷ canvas width × 1000.
Round each result to an integer. For a 1600 × 900 canvas, a rectangle from left 960, top 180 to right 1440, bottom 720 becomes [200, 600, 800, 900]. Notice that vertical coordinates use the height and horizontal coordinates use the width.
The normalized grid stretches with the canvas. Reusing these numbers on a tall portrait image changes the rectangle’s physical proportions. Treat a second aspect ratio as a second layout, not a free resize.
Build variations without losing the composition
For an initial comparison, keep the aspect ratio, object descriptions, and lighting fixed. Change only the mug’s horizontal placement. Compare a centered-right version with a farther-right version and choose the one that leaves sufficient copy space without crowding the handle.
Once a composition is approved, save the image, prompt, element table, and model name. For later image edits, use the documented editing format rather than assuming generation rows are interchangeable with edit rows. BFL uses from, src_bbox, and tgt_bbox to describe the source and destination of an element.
A useful rule is to separate layout decisions from styling decisions. Approve where the objects go first; then test background color or material treatment. Changing placement, lighting, product shape, and camera angle together makes it difficult to identify why a version improved or failed.
Five checks before publishing the image
- Inspect the copy area at its actual display size. Add the real headline in your design tool. Texture that looks subtle at full size may compete with small text on mobile.
- Check geometry. Look closely at the handle, rim, base, and any contact points. Reject duplicated parts or impossible connections.
- Check shadows and overlaps. The mug should rest on the cloth, not float above it. Light direction should agree across the scene.
- Test the destination crop. Preview the real placement. Keep important objects away from areas likely to be cropped or covered by interface controls.
- Verify commercial accuracy. Compare real products against approved references, add only authorized brand assets, and confirm the offer and dates separately from image generation.
Do not interpret a vendor’s precision language as a reason to skip review. BFL’s own tutorial describes boxes as guides rather than hard masks. If an approved asset must remain exactly unchanged, preserve it in a conventional layered composition instead of relying solely on a generative instruction.
Can you use this approach in other AI image generators?
Yes, the planning method transfers: choose a format, divide the canvas into regions, describe each object, and reserve negative space. The exact JSON syntax does not automatically transfer. A generator without documented bounding-box support may treat the coordinates as ordinary text or ignore them.
For a general text-to-image workflow, translate the sample into plain language: “One ivory mug in the right third, handle fully visible, bottom edge near the lower quarter, left half uncluttered for copy.” Evaluate the output against the same checklist.
You can explore text-to-image and image-to-image workflows with AI Photo Generator. Check the currently available model and controls before applying provider-specific syntax. This guide does not imply that FLUX 3 Image or its bounding-box interface is available inside AI Photo Generator.
Frequently asked questions
Do bounding boxes guarantee exact placement?
No. BFL documents them as placement and scale guides, not clipping masks. Leave margins and inspect the output, especially near text areas and crop edges.
Should I generate the headline inside the image?
For a campaign with changing prices, dates, or legal text, editable typography is usually the more controllable choice. If you do generate text, BFL recommends a separate row for each line and the exact wording in its description. Proofread every character.
Is a bounding-box layout the same as an editable design file?
No. A structured prompt describes the intended arrangement; it does not by itself provide a layered design document. Keep your final typography, logos, and other critical assets in a design tool when you need reliable revisions.
The takeaway
The practical value of bounding-box image generation is a more explicit brief. Decide where the subject, supporting objects, and empty space belong before asking for the finished picture. For November campaign work, that turns “make it look better” into a layout you can inspect, revise, and approve.
Sources and verification
Reviewed November 4, 2026. Feature descriptions were cross-checked against BFL documentation and independent reporting; the workflow and sample brief are editorial recommendations, not measured performance claims.
- Black Forest Labs: Bounding boxes with FLUX 3 Image — coordinate order, prompt structure, generation and editing formats, and limitations.
- Black Forest Labs release notes — October 1 FLUX 3 Image release and subsequent changes.
- The Frontier: FLUX 3 Image coverage, October 5 — independent corroboration of layout control and the normalized coordinate grid.
- Teamday: November AI image model comparison, updated October 29 — recent coverage highlighting bounding-box layout control; its broader rankings are not adopted here.