Skip to content
Grok image
English
English简体中文繁體中文DeutschEspañolPortuguês

FLUX 3 Image

A model detail page for FLUX 3 Image, Black Forest Labs' image model released on 1 October 2026. It records what the publisher documents, what third-party reports add, and which of those controls this site's workspace can reach today.

Updated:

FLUX 3 Image

Choose a mode, write the prompt, then read the quoted cost before you generate.

AI Image0

Generated content will appear here

Publisher’s own brief

Three pieces the publisher published with a printed prompt

Each image below is from the publisher's own model page, and each carries the prompt the publisher printed beside it in full. The publisher's remaining gallery pieces demonstrate scale, non-Latin typography and material handling without a printed prompt; they appear in the case list below instead, where captions describe rather than quote.

All three images are Black Forest Labs' own artwork, retrieved on 3 October 2026 from the publisher's FLUX 3 Image page. Rights remain with the publisher; they are shown with source attribution and no claim of ownership. Prompts are reproduced in the language the publisher wrote them in and are not translated, because a prompt is the model input and rewriting one changes what it produces.

Three ballerinas in white tutus dance in formation on a dark stage, arms raised, seen from the wings past the blurred dark silhouettes of two audience members watching. Details and published prompt ↗

01 / Text-to-image, published prompt

Three ballerinas, seen past the audience

Black Forest Labs, FLUX 3 Image model page ↗

Read the published prompt

Three ballerinas dancing Swan Lake, seen from the wings past the dark silhouettes of people watching.

Send this prompt to the workspace ↗

Copied from the publisher’s model page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as callable and priced.

A tall poster of a concrete dome rising from a twilight bay, a crowd seated on the beach before it, with the words LE FESTIVAL DU SOLEIL in a thin serif typeface across the sky. Details and published prompt ↗

02 / Bounding-box layout with text rendering

Le Festival du Soleil, placed by bounding boxes

Black Forest Labs, FLUX 3 Image model page ↗

Read the published prompt

A glowing concrete dome rises from a twilight bay, a crowd on the beach before it.

Send this prompt to the workspace ↗

Copied from the publisher’s model page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as callable and priced.

A fashion portrait of a man in a red beanie, white tee and black knitted vest standing in Times Square, skyscrapers and billboards behind him. Details and published prompt ↗

03 / Multi-reference generation

A streetwear portrait assembled from six references

Black Forest Labs, FLUX 3 Image model page ↗

Read the published prompt

A fashion streetwear portrait in Times Square with ref_image_0, ref_image_1, ref_image_2, ref_image_3, ref_image_4 and ref_image_5.

Send this prompt to the workspace ↗

Copied from the publisher’s model page, where this prompt is printed beside this image. The wording is unchanged, including its original language, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page can accept it: the catalogue returns this product as callable and priced.

A control-first model, launched at the start of October 2026

Black Forest Labs published FLUX 3 Image on 1 October 2026, and the publisher's own one-line pitch is a claim about control rather than beauty: maximum control over every pixel. The release follows FLUX 3 Video, which the same company put online in August, and the image model is the second product to reach users from the FLUX 3 base.

Three capabilities carry that control claim. A bounding-box layout system lets a caller place every element of a composition on a 0-to-1000 grid, each element carrying a stable identifier and a text description. A reference system accepts up to ten input images per generation and cites them in the prompt as ref_image_0 through ref_image_9. And an editing mode regenerates only the boxes a caller marks, holding the rest of the frame in place.

The output ceiling is the other headline. The model generates natively at up to 5456 by 3072 pixels — a 16.8-megapixel frame the publisher calls 4K — without an upscaler in the path. A built-in prompt upsampler expands short instructions into detailed captions before generation, which is a separate feature from the pixel ceiling and easy to conflate with it.

Launch date, the 5456-by-3072 figure and the three capability groups are from the publisher's model page, retrieved 3 October 2026. The 1 October date is corroborated by two independent third-party reports published the following day.

What the publisher demonstrates, and what this workspace exposes

The publisher's page is organised around four interactions, and comparing that list against the fields this site's catalogue publishes is the most useful thing a reader can do here, because one of the four does not reach this workspace at all.

The catalogue column is read from the live site-19 product record on 3 October 2026. The bounding-box system is a structured parameter of the publisher's API, not a prompt convention, so its absence from the form means a visitor here cannot place boxes — not that the underlying service lacks it.

Capability on the publisher pageHow this catalogue exposes itReachable from this site's workspace
Text-to-image from one line and an aspect ratioA text-to-image mode with prompt, aspect ratio and resolution fieldsYes
Bounding-box layout on a 0–1000 gridNot among the published form fieldsNo; layout control cannot be set from this page
Up to ten reference images, cited as ref_image_NA reference-image field capped at 10 on the reference-to-image and image-edit modesYes
Box-by-box editing with untouched regions preservedAn image-edit mode with its own task modelYes, as prompt-guided editing with reference input

Five output tiers, priced per generation

The catalogue returns a five-tier resolution ladder and a per-task point price for each tier, which is more granular than most products in this catalogue. The two lowest tiers cost the same; the top tier is three times the base.

Two further parameters exist in the service schema but render no control on this page. A safety tolerance between 0 and 4 is sent at the service default of 2, and a grounding toggle — which lets the model consult live web sources while writing its prompt — is sent switched on. A visitor cannot change either from this site today.

The publisher's own pricing is a different number and is not what this page quotes. Launch coverage reported an API entry price of about $0.0205 for a 768-pixel image under a fifty-percent launch discount that runs until 8 October 2026, after which the publisher's entry price roughly doubles. This site's workspace quotes in points from its own catalogue, and the two figures are not conversions of each other.

Prices are read from the site-19 product record on 3 October 2026 and are the figures the workspace quotes from before a task starts. The publisher price is third-party launch reporting, quoted as a claim rather than a fact; the publisher does not publish a public price list on the model page itself.

Resolution tierPrice per generationNotes
768sq6 pointsSquare draft tier; useful for composition checks
1K6 pointsThe default tier the workspace opens with
1.5K8 pointsA middle tier most models in this catalogue do not offer
2K12 pointsDoubles the base price
4K18 pointsThe native-resolution tier, up to 5456 by 3072

How the reference and editing prompts are written

The publisher publishes the prompt beside each demonstration, and the reference cases all follow one grammar that is worth learning before typing anything into the workspace.

Each uploaded reference is addressable by a token in the order it was added, starting at ref_image_0. The publisher's own six-reference street-portrait brief reads: A fashion streetwear portrait in Times Square with ref_image_0, ref_image_1, ref_image_2, ref_image_3, ref_image_4 and ref_image_5. Nothing else in the sentence changes — the tokens simply name the images that must appear.

The editing cases use the same token grammar to say what changes and what must not. The publisher's published edit record for its surfer demonstration instructs: change the black wetsuit to a bright red neoprene wetsuit with the same fit and highlights, change the white surfboard to a bright red surfboard maintaining the same shape and reflection, and keep the ocean wave, the cloudy sky and the rest of the composition exactly unchanged. The pattern is a description per changed element plus an explicit keep-list for everything else.

The bounding-box layout system works differently and is the one capability this page cannot reach: it takes a JSON array in which every element carries an identifier, a box on the 0-to-1000 grid and a text description. The publisher's festival example places the headline text, a coastal town, a concrete dome, swimmers and a crowd as five rows of that array. A language model can generate such an array, which is why launch coverage called the layout API the part to watch — but on this site that array has no input field.

The street-portrait prompt and the edit record are quoted from the publisher's model page in the language they were written in; a prompt is a model input and is not translated on this site. The festival element list is the publisher's own five-row example.

What the first wave of independent coverage agrees and disputes

Within two days of launch, the top three independent results for this model's name are a developer-oriented launch analysis, a Chinese-language encyclopaedia entry and a tool-directory listing. They overlap on the facts and diverge on the frame.

The three sources are cited as reporting, not as independent verification: the pixel-preservation percentages originate from the publisher and are repeated by the launch analysis, and no third party has reproduced them. Where a source describes capabilities of the FLUX 3 family rather than of FLUX 3 Image, that distinction is kept.

  • All three agree on the core specification: native output up to about 16.8 megapixels, up to ten reference images, bounding-box layout on a normalised grid, and multi-turn editing that leaves untouched regions in place.
  • The launch analysis adds numbers the others lack: pixel-preservation figures the publisher published for sample edits — 80.7 percent of pixels unchanged when removing a cat from a scene, 89.7 percent when changing a shirt print, 67.8 percent for the harder task of recolouring hair — and the observation that FLUX 2 had slid down text-to-image rankings by August before this release answered.
  • The encyclopaedia entry widens the frame to the whole FLUX 3 base, which it describes as a unified architecture that also generates video with native audio, and places the image model in a staged rollout alongside an open-weights edition that has been promised but not dated.
  • The directory listing contributes the use cases: magazine layouts, product composites, collages, e-commerce scenes and iterative editing — the workloads where placing an element exactly where it belongs matters more than a single heroic render.

The artwork on this page, and what it cannot show

Every image here comes from the publisher's own model page, and every one was opened and looked at before it was used. Nothing on this page is a user submission, a competitor render or a synthetic placeholder.

Screening was visual and per file. Rights remain with Black Forest Labs; images are shown with source attribution and no claim of ownership. Any text visible inside an image — the festival headline, the shop sign, the Japanese poster titles — is part of the generated artwork itself.

  • Fifteen image files are used across the gallery and the cases: six showcase pieces and nine case images, of which two are before-and-after pairs.
  • Prompts are reproduced only where the publisher printed one beside the image. The ballerina brief, the six-reference street portrait and the surfer edit record are quoted in full; the remaining pieces carry captions that describe the image without a prompt, because the publisher does not print one for them in the static page text.
  • One alphabet poster from the publisher's gallery was screened out: it carries a third-party artist credit in the artwork itself, and the rule on this site is to keep no image with a visible third-party wordmark.
  • The bounding-box demonstrations on the publisher page are shown as interactive interface states with overlay graphics rather than as clean stills, so the interface screenshots were not taken as case images.

Examples

Before

Before: A black-and-white photograph of a surfer on a wave, before and after an edit. Before, the wetsuit and board are dark against a grey sea. After, the wetsuit and board are bright red and everything else in the frame is unchanged.

After

After: A black-and-white photograph of a surfer on a wave, before and after an edit. Before, the wetsuit and board are dark against a grey sea. After, the wetsuit and board are bright red and everything else in the frame is unchanged.

The publisher's own recolour demonstration, published as two files. The instruction named the two objects to change and an explicit keep-list — the wave, the sky and the rest of the composition exactly unchanged — and the output holds the surfer's pose, the spray and the light while only the two surfaces turn red. This is the edit record quoted in full in the section above.

Black Forest Labs, FLUX 3 Image model page, retrieved 3 October 2026. The publisher published this edit with its full edit record. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Before

Before: A storybook illustration of cats sharing a roasted mouse by a fire inside an ice cave. Before, three cats sit around the fire. After, the orange tabby on the left is gone and the remaining two cats, the fire and every prop are unchanged.

After

After: A storybook illustration of cats sharing a roasted mouse by a fire inside an ice cave. Before, three cats sit around the fire. After, the orange tabby on the left is gone and the remaining two cats, the fire and every prop are unchanged.

The removal case behind the strongest pixel-preservation figure in launch coverage: the publisher reported 80.7 percent of pixels unchanged when the cat is taken out, and the pair shows what that means — the blanket the removed cat sat on, the fire, the bowls and the snow outside the cave doorway all stay put. The figure originates from the publisher and has not been independently reproduced.

Black Forest Labs, FLUX 3 Image model page, retrieved 3 October 2026. Pixel-preservation figure reported in launch coverage by CodeYourCraft, 2 October 2026. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Spectators at a winter sports event photographed at eye level: a woman in a fur hat and sunglasses holds a Pomeranian, a man beside her holds a Jack Russell terrier, and German and Swiss flags fly in the blurred background.

A crowd study that has to keep two dogs, two coats and two flags honest at once. It is the kind of frame where a generator usually invents a third dog or mirrors a flag incorrectly; the publisher includes it among its gallery pieces without a printed prompt, so this caption describes rather than quotes.

Black Forest Labs, FLUX 3 Image model page, retrieved 3 October 2026. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Five landscape photographs pinned on a white ground: a sunlit mountain with a cloud cap, an aurora over a silhouetted couple, a stone cottage in snow, a city skyline across a river, and the blue interior of an ice cave.

Five frames of winter in one composition — the publisher's title for the piece. A collage tests whether the model can hold five unrelated scenes in one layout without letting any of them bleed into another, and the white gutters between the photos stay clean.

Black Forest Labs, FLUX 3 Image model page, retrieved 3 October 2026. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Knitted iceberg peaks with tiny knitted houses under a knitted banner reading KNITTED ICEBERG VILLAGE, a pink whale tail rising from tangled teal yarn.

The same piece shown in the gallery above, kept in the case list because it earns its place twice: the lettering in the banner is knitted, not printed, which makes it a harder typography test than a flat sign, and the yarn sea tangles into dozens of separate loops without repeating itself.

Black Forest Labs, FLUX 3 Image model page, retrieved 3 October 2026. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A pop-art poster of a purple demon girl in a red hood against yellow, with vertical Japanese lettering on both sides and a red seal.

The Japanese-script case. Independent coverage of the launch singled out the publisher's Japanese examples, and this poster sets two vertical writing columns plus a horizontal line and a seal mark, all of which have to read correctly for the poster to work.

Black Forest Labs, FLUX 3 Image model page, retrieved 3 October 2026. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A night exterior of a small Japanese soba restaurant with an illuminated sign, red lanterns, a poster of noodle dishes in the window and a bicycle leaning outside.

The native-4K demonstration. The publisher presents this frame at 5456 by 3072 — 16.8 megapixels generated in one pass — and zoomable crops of the shop sign, a lantern and a diner are published beside it to show the detail survives enlargement. What is shown here is the full frame scaled down.

Black Forest Labs, FLUX 3 Image model page, retrieved 3 October 2026. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Three ballerinas in white tutus dancing in formation on a dark stage, seen past the silhouettes of two audience members.

The publisher's one-line text-to-image brief, shown here as the entry point of the whole system: one sentence and an aspect ratio. The silhouettes in the foreground are part of the prompt, not an accident of framing, which is what makes the brief a good test of whether the model places what is asked for.

Black Forest Labs, FLUX 3 Image model page, retrieved 3 October 2026. The publisher published this image with the prompt shown beside it. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Questions

What is FLUX 3 Image?

It is the image generation and editing model of Black Forest Labs' FLUX 3 family, released on 1 October 2026. Its defining features are layout control — every element of a composition can be placed with a bounding box on a 0-to-1000 grid — up to ten reference images cited in the prompt as ref_image tokens, multi-turn editing that leaves untouched regions unchanged, and native output up to 5456 by 3072 pixels without an upscaler.

Can I generate images with it on this site?

Yes. The site-19 catalogue returns FLUX 3 Image as callable and priced in all four task types — text-to-image, image-to-image, reference-to-image and image-edit — so the workspace on this page can send a task. It pins the model when the page opens and the server quotes the point price before the task starts.

How much does a generation cost here?

The catalogue prices each generation in points by resolution tier: 6 points at 768sq or 1K, 8 points at 1.5K, 12 points at 2K and 18 points at 4K. That is the figure the workspace quotes from before a task runs. The publisher's own API price is a different number reported in launch coverage and is not what this site charges.

Can I use the bounding-box layout feature here?

No. Bounding-box layout is a structured parameter of the publisher's API, and it is not among the fields this site's catalogue publishes for the product, so there is no input for it on this page. What you can do here is text-to-image, reference-based generation with up to ten images, and prompt-guided editing. The catalogue column in the table above records exactly which capabilities the form exposes.

How do reference images work in the prompt?

Each uploaded image gets a token by its position in the upload order, starting at ref_image_0. You cite the tokens in the prompt text — for example, a six-image street portrait brief reads A fashion streetwear portrait in Times Square with ref_image_0 through ref_image_5. For edits, the same grammar names what to change and includes an explicit keep-list for everything that must stay unchanged.

What resolutions can I choose, and which is the default?

Five tiers: 768sq, 1K, 1.5K, 2K and 4K, with 1K as the default the workspace opens with. The 768sq draft tier and the 1.5K middle tier are two options most other models in this catalogue do not offer. The top tier is the native-resolution mode the publisher demonstrates at 5456 by 3072.

Is FLUX 3 Image open source?

Not yet. The model is currently offered as a paid service, and both the publisher and independent reports say an open-weights edition is planned for the coming weeks without a firm date. Commercial weight licensing is handled by contacting the publisher directly. When an open edition ships, it will be a separate product from the API this site calls.

Why are the prompts on this page not translated?

Because a prompt is a model input, not a sentence for a reader. The publisher wrote these briefs in English, and the Japanese poster case depends on Japanese text inside the image, not in the prompt. Translating a prompt replaces the text that produced the image with a paraphrase, so every language version of this page shows the same prompts in their original wording.

Explore next