FLUX 3 Image: How BFL’s Layout Controls Fix AI Ad Creative

Every marketer who has used AI image models knows this moment: the client approves the layout, then asks for one change — the bottle in green, a different headline, another model. On most image models that means re-rolling the entire composition and hoping the other 90% lands the same way twice. Black Forest Labs released FLUX 3 Image on October 1 with a different proposition: tell the model where every element belongs, then change one element while the rest of the composition stays put.

This is less "another prettier image model" and more a new control surface for the part of creative work that is actually expensive — layout and revisions. Here is what launched, what it costs, and which claims still need your own testing.

What actually launched on October 1

FLUX 3 Image is the image half of Black Forest Labs’ FLUX 3 family, the multimodal system announced in July that also covers video and audio. It is live now through the BFL API and a browser Playground, with commercial weights available by license and public open weights promised in the coming weeks. Generation and editing both run through a single endpoint.

• Up to 10 reference images per generation, supplied as URLs or base64.

• Native high-resolution output — five tiers from 768×768 to 4K (about 16 megapixels; 1K is the default).

• 15 aspect ratios, from 21:9 to 9:21, plus an auto mode that takes framing from your prompt or first reference.

• A layout system built on named elements and bounding boxes — the headline feature for ad work.

• Element-level editing that treats an image as parts you can keep, move, add or remove individually.

The layout system: named elements on a 0–1000 grid

A normal prompt — "bottle on the right, headline at the top, copy on the left" — leaves the geometry to the model’s mood. FLUX 3 Image replaces that with structured input: a one-line caption describing the scene, followed by a JSON array of elements. Each element carries an id, a bbox and a desc, and the caption cites each element by its id.

The canvas is a 0–1000 grid on both axes at any aspect ratio, and each box is written [y_min, x_min, y_max, x_max]. Because the coordinates are normalized, the same layout works at 1K for exploration and at 4K for the final master.

Here is a complete layout prompt for a skincare ad, in the documented format:

caption = ("A minimalist skincare ad. The headline <headline_1> sits above "
           "campaign copy <copy_1>. A serum bottle <product_1> stands right.")

elements = [
    {"id": "headline_1", "bbox": [40, 80, 190, 920],
     "desc": "Text reading ’Wake up brighter’, charcoal serif"},
    {"id": "copy_1",     "bbox": [240, 80, 520, 460],
     "desc": "Text reading ’Vitamin C serum. Cold-pressed.’"},
    {"id": "product_1",  "bbox": [230, 560, 900, 900],
     "desc": "Frosted glass serum bottle, amber liquid, soft studio light"},
]

payload = {
    "prompt": caption + " " + json.dumps(elements),
    "aspect_ratio": "4:5",
    "resolution": "1k",
}

Three things make this more than syntax sugar:

• The layout is inspectable. A script can check that every element has a box, that coordinates stay on the canvas, and that the caption mentions every id — a design brief you can lint.

• It is versionable. "Move the product up 80 units" is a one-line change to a stored spec, not a freshly written paragraph of persuasion. That is what surviving client revisions looks like as data.

• Your boxes are respected. BFL’s prompt upsampler expands your short caption into the dense prose the model was trained on, but every box you drew reaches the model verbatim, with the same id and the same coordinates.

You can draw boxes by hand in the Playground or send the JSON via the API — the boxes travel inside the prompt string itself; a separate entities field returns a validation error. One honest caveat from BFL’s own docs: boxes guide placement and scale, but elements may extend beyond them. If you need hard safe areas — legal lines, packaging marks, an approved face — verify them in the output rather than trusting the box as a wall.

Edits that change one element, not the whole brief

Editing reuses the same element table with four operations: keep (an element keeps its position), move (same element, new box), new (generated in a target box) and remove (target box set to null). Several edits can go into one request, and edits chain across turns.

BFL markets this as pixel-perfect editing. Its documentation is more careful: pixels outside the edited boxes usually stay identical. That word is doing real work. Recolouring a bottle changes its reflections; removing a prop reveals background that has to be reconstructed; the seam around an edit is exactly where artefacts show up. Treat it like a retoucher: check the edited region and its boundary, and keep the previous version alongside every revision.

For client work, the structure is the point. "Change the label to gold" and "move the bottle left" are different jobs, and FLUX 3 stores them differently — one changes an element’s description, the other its coordinates. The rest of the approved layout is preserved by design rather than by luck.

Up to 10 references, tokenized

Reference images get tokens in the order you add them — ref_image_0, ref_image_1, and so on — and one prompt line cites them all. FLUX 3 decides where each reference sits and how big it renders. Limits: at least 256×256 pixels, at most 16 megapixels per reference, 20MB per base64 payload. Note that BFL downscales references above 4MP before processing, so uploading a giant product photo does not mean the model sees it at full resolution.

For e-commerce creative this is the "everything in one frame" case: product shot, packaging, the model from the last campaign and the set design, combined in one composition without a composite tool.

Pricing: cheap until October 8

BFL is halving image rates for requests accepted between October 1 and October 8, 8am Pacific (15:00 UTC) both days. References and prompts are included in the per-image rate, and failed or moderation-blocked generations are not charged.

The jump that matters is 2K to 4K: $0.10 to $0.607 at list prices, a 6.07× step. The sensible strategy is exploring at 1K and mastering at 4K — with one caveat: a 4K request is a new generation, not an upscale of the approved 1K draft, so do not promise the client an identical image at the end.

The iteration math is friendly. Four 1K candidates, two 1K revisions and one 4K master cost about $0.45 during launch week, $0.90 at list prices. Cheap generations only stay cheap if you do not have to re-roll — which is precisely what the edit controls are for.

What nobody knows yet

FLUX 3 Image had no entry on the public leaderboards on launch day: the Artificial Analysis text-to-image and image-editing boards listed no FLUX 3 result when checked on October 1. The current field there is led by GPT Image 2.5 Sunburst (Elo 1197) and Flare (1190), ahead of Nano Banana 2 (1125), Seedream 5.0 Pro (1081) and FLUX.2 [max] (1021). An absent score is not a bad score — but it means nobody can yet tell you FLUX 3 beats those models on raw quality or edit reliability, and FLUX 3 Video’s strong preference results do not transfer to still images. Kingy.ai’s launch analysis has the full breakdown.

What to test before you trust it with client work:

1. A fixed ad layout. Product, headline and copy in predefined regions — did every element land inside its box?

2. Real typography. Render your actual headline and body copy, then change one line and check the untouched text did not drift.

3. A five-edit sequence. Recolour, replace, move, add and remove on one approved image, comparing protected regions after each step.

4. Ten references at once. One composition, ten inputs — count the missing and duplicated objects.

5. Matched briefs at 1K and 4K. Same layout, both sizes. Check latency too: BFL warns 4K generations can take minutes, and every job is asynchronous.

Should ad teams switch?

If your workflow is "describe it once and hope," the current leaders — GPT Image 2.5, Nano Banana 2 — remain strong, benchmark-proven choices. If your workflow is "approved layout, then surgical revisions," FLUX 3 Image is the first major model whose interface was built for exactly that job. The structured layout turns a design into data you can version, lint and hand to a client, and the element-level edits target the real cost of creative production: not the first draft, the fifth revision.

There are two ways in: draw boxes by hand in the BFL Playground, or send layout prompts to the API. If you go the API route, poll the returned job URL and download the signed result within an hour, and keep your API key server-side — never in a client app.

Try FLUX 3 on Automatica

FLUX 3 is live on Automatica now — detailed, well-composed images with clean text rendering at up to 4K, with FLUX 3 editing alongside it. Generate and iterate in a clean, fast interface and pay for what you create.

Create with FLUX 3 on Automatica →