Mis à jour
What is FLUX 3 Image?
FLUX 3 Image (the flux 3 image API) is Black Forest Labs' full image model from the FLUX 3 family, released on October 1, 2026. FLUX 3 is one model trained on images, video and audio together, and this is the part that makes and edits stills. Text-to-image, editing and multi-reference generation all run on one route.
What sets it apart is layout control. You can append bounding boxes to the prompt, one row per element, and the model composes the image around them. The same boxes drive edits: recolor, move or replace one element and leave the rest of the frame where it was. No mask, no separate inpainting call.
On UnificAlly the route takes a prompt, up to ten reference images, auto plus 15 aspect ratios from 21:9 to 9:21, five resolution presets from 768sq to 4k, and a grounding switch that lets the model search the web before it draws. Billing is per finished image by resolution. If you want open weights and a smaller model, FLUX 3 Image Klein is the 5.6B sibling.
Key features of FLUX 3 Image
Bounding boxes in the prompt
Each element gets an id, a box on a 0 to 1000 grid and a short description. The headline lands in its box, the caption in its corner, the subject where you put it. An app can generate these rows, so layouts stop being guesswork.
Up to 10 reference images
Send up to ten images in image_urls and say what each one supplies: the person, the product, the setting, the style. With one image it is an edit; with several it is a composite.
Text that comes out spelled right
Headlines, dates and small print render as written, in the type style you ask for. Good for posters, packaging comps and ad layouts that carry real copy.
Five resolutions up to native 4K
768sq, 1k, 1.5k, 2k and 4k. Each is a fixed pixel area, so the exact size follows the aspect ratio. 4k renders at about 16 MP rather than upscaling a smaller image.
Web grounding, on by default
With grounding on, the model searches the web and images for the real things in your prompt before it draws: a landmark, a product, this year's event poster. Switch it off for invented subjects and faster runs.
What FLUX 3 Image is best for
Posters, ads and editorial layouts
Put the headline, the photo and the caption in their own boxes. Type set around a picture is the case bounding boxes were built for.
Catalog and product edits
Recolor a product, swap a background or move one object with box-level edits. The parts you did not touch stay put, so the image still matches the rest of the set.
Multi-reference composites
Combine a person, a product and a scene from up to ten references in one call, with the prompt saying which image supplies what.
Agent-built images
An LLM can write the caption and the element table from a one-line brief. The layout becomes data your app controls, and every box stays editable.
Real-world subjects
Grounding pulls in web and image search, which helps with named places, products and events that a model would otherwise guess at.
Print and large-format work
Find the composition at 1k, then rerun the approved prompt at 2k or 4k for print, hero banners or crops you need to zoom into.
FLUX 3 Image use cases
Build a poster or social-ad generator where the headline, date and venue come from form fields and each one lands in a fixed box, spelled right. Run a catalog tool that takes the existing product shot as a reference and recolors or restages one element per request without regenerating the whole frame. Set up a lookbook pipeline that places a model, a garment and a location from separate references into one composite. For an editorial app, let an LLM turn a short brief into a caption plus element table, show the boxes to the user for a quick drag, then render at 2k.
FLUX 3 Image limitations
There is no seed, so the same request can return a different image; lock a look by editing an approved image instead. Boxes guide placement but are not clipping masks, and shadows, reflections or nearby lighting can still change outside an edited box. Very small boxes often fail: Black Forest Labs reports that a new element in a box of about 40 x 25 pixels often does not appear, so give new content room. The grid also stretches with the frame, so a box that is square at 1:1 becomes a wide rectangle at 21:9. Grounding adds a search step, which makes runs slower than with it off.
What FLUX 3 Image can do
House-generated examples on this route.
A layout drawn with bounding boxes
Five boxes in the prompt: a serif headline across the sky, pines on the right, a black cabin in the middle, a red blanket on its porch rail and a small caption in the lower-right corner. Every element sits in its box, and both text lines are spelled as written. This is the control you do not get from a plain caption.
Two references, one scene
Image 1 is a portrait, image 2 is a macro of a watch movement. The prompt puts the woman at a watchmaker's bench with that movement in front of her. Her face, freckles and silver-streaked curls carry over, and the movement keeps its blued screws and red jewels.
Poster type, every line correct
A two-color riso poster with a big condensed headline and three lines of small print. "BLUE HOUR SESSIONS", the date, the venue and the ticket line all come out as written, with the paper grain and slight misregistration the prompt asked for.
Photoreal detail at 2k
A 2512x1664 documentary shot from the 2k preset. The knots in the orange net, the loose threads on the sweater and the salt on the boots all hold at full size, and the dawn light reads as one consistent source across the pier and the water.
FLUX 3 Image vs FLUX 3 Image Klein
Both come from FLUX 3 and both read bounding boxes. Klein is the 5.6B open-weight model: on our route it takes four references, seven aspect ratios up to 4 MP, a steps control and a seed, which makes it the one for repeatable batches and for teams that also run weights locally. FLUX 3 Image is the full model: ten references, auto plus 15 aspect ratios, a 4k preset at about 16 MP and web grounding, with no seed. Use Klein for fast, seeded, high-volume work, and FLUX 3 Image when the layout, the reference count or real-world accuracy matters more. Arena.ai had no FLUX 3 Image entry on its text-to-image or image-edit boards as of October 22, 2026, so there is no independent rank yet.
FLUX 3 Image API examples
Call FLUX 3 Image from any language by POSTing to /v1/tasks. Full parameter docs live at docs.unifically.com/models/image/flux/flux-3-image.
curl -X POST https://api.unifically.com/v1/tasks \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "black-forest-labs/flux-3-image",
"input": {
"prompt": "A screen-printed concert poster with the headline \"BLUE HOUR SESSIONS\" above an upright bass player on cobalt blue",
"aspect_ratio": "2:3",
"resolution": "1k",
"grounding": false
}
}'
Successful submission returns a task_id. Poll GET /v1/tasks/<task_id> or set a callback_url on the request to receive the finished image URL.
FAQ
Questions fréquentes sur FLUX 3 Image
FLUX 3 Image is the image side of Black Forest Labs' FLUX 3 model, released on October 1, 2026. One model does text-to-image, editing and multi-reference generation, takes bounding boxes inside the prompt to place each element, and renders up to native 4K. On UnificAlly it is the black-forest-labs/flux-3-image model.
A prompt, up to ten reference images in image_urls, aspect_ratio (auto plus 15 fixed ratios from 21:9 to 9:21), resolution (768sq, 1k, 1.5k, 2k or 4k, default 1k) and grounding (on by default). There is no seed, steps or guidance setting; control lives in the prompt.
You write a normal caption, tag each element with an id like <cabin_1>, then add a JSON array where each row has an id, a bbox and a desc. Boxes are [top, left, bottom, right] on a 0 to 1000 grid at any aspect ratio, so [0, 0, 500, 500] is always the top-left quarter. The same rows, with source and target boxes, drive edits that move, replace or recolor one element.
With grounding on (the default), the model runs a web and image search on the prompt before it draws, so real landmarks, products and events come out closer to the real thing. Turn it off for invented subjects or when you want a faster run that uses only your prompt.
Up to ten in image_urls, each between 256 x 256 pixels and 16 megapixels. Refer to them in the prompt as image 1, image 2 and so on, and say what each one supplies. With aspect_ratio on auto, the output keeps the framing of the first reference.
Five presets, each a fixed pixel area whose exact size depends on the aspect ratio. 1k is about 1 MP (1360 x 768 at 16:9), 2k is about 4 MP (2512 x 1664 at 3:2), and 4k is about 16 MP. 768sq and 1.5k sit below and between them.
Not with a seed. FLUX 3 Image has no seed parameter, so the same request can return a different image. To keep a look fixed, pass the approved image back as a reference and edit it with bounding boxes, which leaves the untouched areas in place.
Per finished image, by resolution preset. Reference images do not add to the price, and failed generations are not charged. The current rates are on the pricing page.
