FLUX.3 Image Prompt Guide: Text, Layout Boxes and 4K

All guides · FLUX.3 Image in Fuser

Quick answer: describe the medium, subject, setting, light and framing in plain sentences, and set the aspect ratio yourself. Put every word you want on the image in double quotes and say where it goes and what the type looks like. Tie each hex code to one named object. Describe what should be there instead of what should not, because FLUX.3 Image has no negative prompt. For exact placement, end the prompt with a JSON list of bounding boxes. Draft at 768px or 1K and save 2K and 4K for finals. In our tests a five-block multilingual poster came back with every string spelled correctly, and a box layout put the masthead and issue line inside their boxes, all from the Prompt field of the FLUX.3 Image node.

What FLUX.3 Image is

FLUX.3 Image is Black Forest Labs' newest image model, released through its API on 1 October 2026 (release notes). It is the image side of FLUX 3, which BFL announced in July as one model trained on images, video and audio, saying it can "render high-accuracy text in multiple languages" (announcement). BFL ships it as a single endpoint where the prompt decides whether to generate or edit, with up to ten reference images, bounding boxes written into the prompt and output up to 4K, which BFL puts at about 16 megapixels (release notes).

In Fuser it is one node, FLUX.3 Image. With no images connected it generates from text; connect between one and ten images and the same node edits or combines them (covered in the FLUX.3 Image editing guide). The node exposes:

• Prompt (required), a text field you can also feed from a Text node or a chat node.

• Images, zero to ten references. With Aspect Ratio on Auto, the first one sets the frame unless the prompt asks for a ratio.

• Resolution: 512px, 768px, 1K (default), 2K or 4K.

• Aspect Ratio: Auto, 21:9, 2:1, 16:9, 3:2, 7:5, 4:3, 5:4, 1:1, 4:5, 3:4, 5:7, 2:3, 9:16 or 1:2.

• Expand Prompt, off by default.

• Output Format: JPEG (default) or PNG.

There is no seed and no negative prompt. That matches BFL's API, which rejects unknown fields such as seed with an error (text to image) and "has no negative_prompt or prompt_upsampling field" (technical parameters). Every run is a fresh sample, so judge a prompt on more than one image before you rewrite it. Not to be confused with FLUX.3 Video, a separate node.

Structure a text-to-image prompt

BFL's advice is to "describe the subject, setting, light, and framing" (text to image), to open with the medium and look, and to name the camera, lens or film stock instead of writing "professional photo" (style guide). For light it asks for the source, direction and quality: "soft window light from the left" says more than "good lighting".

We tested the difference with a four-word prompt and a detailed one, at 3:2 and 1K, one run each:

All three images are coherent night harbours, but only the detailed prompt gave us the scene we asked for: the green-and-white ferry on the left, the striped kiosk on the right, cobblestones reflecting orange light and a lighthouse on the horizon. The short prompt let the model choose everything, and it chose a car-ferry terminal with its own signage. Write the prompt in this order:

1. Medium and look: "A wide documentary photograph", "a flat, screen-printed poster".

2. Main subject, with position: "A single green-and-white ferry is moored on the left".

3. Setting and secondary elements, each with a place in the frame.

4. Light: source, direction, colour.

5. Capture: camera, lens or film stock, and what it does to the image ("visible grain, cool blue shadows").

Set the aspect ratio, or write it in the prompt

BFL's prompting guide says Auto with no reference images "produces 1:1, so set the ratio yourself for any text-to-image request that is not square" (technical parameters). Its API reference says that under Auto "a ratio the prompt asks for wins" and, without references, "the prompt decides, else 1:1" (API reference). Our runs matched the API reference. The detailed ferry prompt at Auto, which says "wide" but names no ratio, came back 1360 × 768 rather than square. The poster prompt below opens with "Create a finished 2:3 event poster" and came back 832 × 1248, exactly 2:3, with Aspect Ratio on Auto.

Because the two BFL pages disagree, the safe habit is to pick the ratio on the node. If you leave Auto on, state the ratio in the first sentence of the prompt, as BFL does in its own poster examples. Either way, make the prompt use the frame: "cliffs on one side, turquoise sea on the other" fills a 21:9 frame that a single centred subject would leave empty.

Put exact text in the image

For each block of text, BFL asks for three things: the exact string in quotation marks with the capitalisation and punctuation you want, where it sits in the frame, and how it looks (weight, case, colour). Inside a quoted string, \n marks a line break, and "arranged vertically" sets the reading direction (text in images).

We wrote one poster prompt (sent as one paragraph; shown here a sentence per line) with five text blocks in three scripts: an English title, a French subtitle with accents, vertical Japanese, a date line with an en dash and middots, and a two-line footer split with \n.

Create a finished 2:3 event poster for an invented harbour festival.
A night harbour scene in a flat, screen-printed style: deep navy water (#1B2A49), a small green-and-white ferry outlined in cream, and strings of warm round lanterns crossing the sky.
At the top, the title "HARBOUR LIGHTS" in huge bold condensed sans-serif capitals, cream, spanning the full width.
Directly below the title, a smaller line in a thin italic serif reading "Fête des lumières du port".
Down the right edge, the Japanese words "港の光" in red brush lettering, arranged vertically.
At the bottom centre, a line of small white sans-serif text reading "12–14 June · Pier 4 · 19:00".
In the bottom-left corner, two lines of small cream text reading "Free entry\nAll ages welcome".

In one run, every string came back spelled exactly as quoted, including "Fête", "lumières", the Japanese characters set vertically in red brush lettering, the en dash and both middots. The footer broke onto two lines where we put \n. The one miss was placement: we asked for the date line at the bottom centre, and the model moved it right to share the bottom band with the footer. That is one run of one prompt, so treat it as a good sign rather than a guarantee, and do what BFL says: proofread every quoted string before you publish, especially small print.

For posters with many blocks, BFL recommends giving each block its own bounding box (next section). For a comparison of text accuracy across models, see the best AI models for text in images.

Use hex codes for exact colours

A hex code names an exact colour, and BFL's rule is to "tie each hex code to one named object" so the model knows where the colour goes (style guide). We tied four codes to four objects in one product shot:

A studio product photograph of three stacked ceramic bowls on a seamless paper backdrop.
The top bowl is glazed burnt orange #E4572E, the middle bowl is deep navy #29335C, and the bottom bowl is saffron yellow #F3A712.
The backdrop and the surface beneath are warm off-white #EDE7DC.
Soft window light from the left, a gentle shadow falling to the right, matte glaze with fine speckles, shot on a 90mm lens at f/8.

Each colour landed on the right bowl, in the right order. The sampled values sit close to the requested hues but are all less saturated, and the light shifted their brightness: the orange, yellow and backdrop came back darker, the navy lighter. Hex codes get you the right colour family on the right object. If a brand colour has to be exact, check it in the output and correct it downstream rather than expecting the photograph to reproduce the swatch.

Describe what you want, not what you don't

Without a negative prompt field, BFL's advice is to "describe the intended scene directly: 'an empty promenade' instead of 'no crowds'", and it lists swaps such as "empty", "deserted" or "solitary" for "no people" (technical parameters). We ran the two wordings once each:

The prompt that said "no people, no cars, no birds" produced a fine square with a few small figures at the far end of the street. The prompt that said "empty, deserted" produced nobody at all. One pair is not proof, but it matches BFL's guidance, and the fix costs nothing: name what fills the space ("bare wet cobblestones", "chairs stacked on the café tables") instead of what is missing.

Control the layout with bounding boxes

This is FLUX.3 Image's distinctive feature, and it needs no extra setting. A layout prompt is a caption that names each element with a tag such as <title_1>, then a space, then a JSON array with one row per element: an id, a bbox and a desc (layout prompts). Every box is [top, left, bottom, right] in whole numbers from 0 to 1000, measured from the top-left corner, so [0, 0, 500, 500] is the top-left quarter at any size or ratio (bounding boxes). Quote the words for a text element inside its desc.

Because the boxes live in the prompt, they go straight into the Prompt field of the FLUX.3 Image node. We pasted this layout as one line (caption, a space, then the array; it is broken into lines here for reading), set Aspect Ratio to 2:3 and left Expand Prompt off:

A flat, printed magazine cover in a clean editorial style, photographed straight on.
A cream paper background <background_1> fills the page.
The masthead <title_1> runs across the top.
A vintage red bicycle <bicycle_1> leans at an angle in the lower left of the cover, its front wheel cut off by the left edge.
A short cover line <subtitle_1> sits in the upper right, and the issue details <issue_1> sit in the bottom-right corner.
[
 {"id": "background_1", "bbox": [0, 0, 1000, 1000], "desc": "Warm cream paper (#F1E8D8) with a faint fibrous texture."},
 {"id": "title_1", "bbox": [30, 60, 180, 940], "desc": "Huge bold black condensed sans-serif capitals reading \"SLOW ROADS\", filling the width."},
 {"id": "bicycle_1", "bbox": [420, 0, 960, 640], "desc": "A vintage steel road bicycle painted cherry red, with cream bar tape, a brown leather saddle and a small wicker basket on the front, in soft daylight with a gentle shadow."},
 {"id": "subtitle_1", "bbox": [230, 560, 400, 940], "desc": "Right-aligned black serif text in two lines reading \"The long way\nround Lake Geneva\"."},
 {"id": "issue_1", "bbox": [880, 700, 960, 940], "desc": "Small black sans-serif text reading \"No. 14 · Autumn 2026\"."}
]

For the left image we sent the caption alone. The caption already describes positions in words, so the composition came out similar, but the model invented its own masthead, cover lines and issue details. With the element table, the cover used our exact words, our type descriptions and the bicycle as described, with the basket and cream bar tape. Drawing the boxes on the result shows how closely it held:

• The masthead and the issue line sat inside their boxes.

• The subtitle kept its right edge and line break, but at that type size it ran left past its box.

• The bicycle filled the lower left, with the front wheel running past the right side of its box. Our caption said the front wheel would be cut off by the left edge; the model cut off the rear wheel instead.

• The issue line "No. 14 · Autumn 2026" was split onto two lines, and the middot turned into a bullet at the start of the second line.

That is what BFL documents: "Boxes guide placement; they are not clipping masks", and an element can extend past its box (bounding boxes). BFL lists a few more tips. Set the aspect ratio you drew the boxes for, because the 0 to 1000 grid stretches with the frame. Give each line of text its own row. Do not make boxes too small: BFL says a new element in a box of about 40 × 25 pixels often failed to appear. If the exact line breaks matter, size text boxes generously and put the break in the quoted string.

Writing coordinates by hand is slow. BFL suggests letting a language model draft the caption and element table from a one-line idea (bounding boxes). On a Fuser canvas you can do that with a chat node such as Claude Chat and connect its response to the FLUX.3 Image Prompt input, then adjust the numbers before you run it.

Pick a resolution

Resolution tiers are equal-area classes, so exact dimensions change with the aspect ratio (API reference). BFL's advice is to iterate at a low tier and use 2K or 4K for detail, adding that "changing the resolution makes a new generation, so the composition can differ from the draft" (technical parameters). We ran the detailed ferry prompt at 3:2 on each tier:

The prompt was specific enough that every tier kept the same arrangement, but the ferry, kiosk and lamp posts were redrawn each time. 4K returned 5024 × 3344, about 16.8 megapixels, in line with BFL's "about 16 megapixels". In Fuser, 2K costs about twice a 1K image and 4K more than twelve times; 768px costs a little less than 1K. The node also lists a 512px tier, but our requests at 512px were rejected with an error at both 3:2 and 1:1, and BFL's own API lists 768 as its smallest tier, so draft at 768px. Lock the prompt at 768px or 1K, then render the final at 2K or 4K, and expect the final to be a sibling of the draft rather than an enlargement of it.

Expand Prompt

Expand Prompt is off by default in Fuser and, per the node's description, expands the prompt "while preserving its intent and the role of each reference image". BFL's text-to-image page says a finished result includes "the expanded prompt the model used" (text to image). The Fuser node returns only the image, so you cannot read any rewritten prompt on the canvas.

Our one pair with the four-word ferry prompt, off and on, gave two different but equally plausible terminals. Without a seed we cannot say which differences came from the toggle and which from sampling. We left it off for every other test so each prompt reached the model as written, and we suggest the same for text and layout prompts, where you have already decided what the image should contain.

Where it fits in a Fuser workflow

A FLUX.3 Image node is one step on a canvas. Feed its prompt from a Text node you can reuse across runs, connect the output to a second FLUX.3 Image node with an image input to fix one detail (see the editing guide), or pass it to a video node to animate it. For an exact pixel size, such as a banner, generate at the nearest ratio and crop in the Compositor. For how it compares with other models, read FLUX.3 vs FLUX.2, FLUX.3 Image vs GPT Image and, for the previous generation's prompting, the FLUX.2 prompt guide.

All outputs in this guide were generated on 2 October 2026 with the same model version and settings the Fuser node uses, one run per prompt, with no reruns or selection.