Describe the picture you want, attach up to ten references, or mark the exact place an element belongs. FLUX 3 Image is Black Forest Labs' current image model, and this workspace runs the four modes it publishes.
These are Black Forest Labs' own layout samples. Each image was generated from the prompt printed with it, and that prompt is the part worth reading: it names one box per element and then holds the rest of the frame still.
Official Black Forest Labs material. None of these images was generated on this site, and none of them shows a run that this workspace performed.
Studio photograph of a fashion mood board arranged as a grid of fifteen rectangular images separated by clean white borders <board_1>. The top row contains five images: a woman sitting with a wide stance in a black leather jumpsuit <image_1>, a close-up of legs in red tights <image_2>, a blurred figure in a dark blue velvet garment <image_3>, a woman in a black ribbed turtleneck with closed eyes <image_4>, and a black-and-white shot of legs in high heels on a chair <image_5>. The middle row features a high-contrast face close-up <image_6>, a portrait of a woman with closed eyes and a pearl earring <image_7>, the back of a head with wavy hair being combed <image_8>, a motion-blurred facial profile <image_9>, and a figure mid-stride in a dark skirt and boots <image_10>. The bottom row displays a black-and-white profile of a bob haircut <image_11>, an extreme close-up of an eye and damp hair <image_12>, white flower petals on a black background <image_13>, a low-angle shot of a teal leather jacket and silver trousers <image_14>, and a close-up of a hand wearing silver rings <image_15>.
A stylized digital illustration in a flat, graphic anime style, resembling a modern pop-art poster with non-directional, uniform lighting. The composition is built on a vibrant background <background_1>. Framed in a medium close-up from the chest up, a young demonic woman <character_1> occupies the center, angled in a three-quarter view and looking back over her shoulder. Large, bold Japanese typography <Ja_Text_1> runs down the left side of the frame, while another block of typography <Ja_Text_2> is positioned in the upper right section. In the lower right corner rests a square seal <Ja_Text_3>. At the bottom center, a line of bright text <Ja_Text_4> sits directly above a smaller line of text <Unknown_Text_1>.
High-angle black and white photograph with a grainy film texture. In the right foreground, a brightly lit, pale, flat rooftop <rooftop_1> anchors the composition. Four men in dark suits <person_1> <person_2> <person_3> <person_4> walk in mid-stride towards the right edge of the building, rendered almost entirely as silhouettes against the stark surface. They cast long, sharp diagonal shadows that stretch out before them under harsh, directional lighting. To the left, a plunging view reveals a deep urban canyon and a congested city street <street_1> far below, bisecting the dense architecture. The avenue is packed with a multitude of cars and buses <vehicle_collection_1> appearing as small dark shapes, alongside tiny figures of pedestrians <pedestrian_collection_1> dotting the sidewalks and crosswalks. The background is filled with towering, monolithic skyscrapers <skyscraper_collection_1> featuring repetitive grids of windows.
Vertical 35mm photograph with fine grain and slightly soft focus. Centered at the top is a cream-colored title <Fr_Text_1>. A hazy, pale grey and green overcast sky <sky_1> sits above dark, rolling mountains <mountains_1> and a distant coastal town <town_1> on the right. In the dark water <water_1> of a bay stands a massive, pale concrete parabolic dome <dome_1>. The dome's arched opening faces the shore, radiating a warm, golden-orange glow from its interior. A dark rectangular monolith <monolith_1> sits deep inside at water level. The golden light reflects on the rippling water, where scattered silhouettes of people <swimmers_1> wade. In the foreground, a pale, pebbly beach <beach_1> holds a large crowd of seated onlookers <crowd_1> in casual, light-colored clothing.
Full-color architectural photograph shot on 35mm film from a low-angle perspective at ground level, featuring bright, direct sunlight and visible film grain. Under a clear, pale blue sky <sky_1>, a Brutalist concrete structure <building_1> dominates the right and center of the frame. The building features three clusters of tall, nested parabolic arches. At the base of the central section, a flat concrete overhang <overhang_1> shelters a small, dark entrance <entrance_1>. Mounted on the overhang is a large, rectangular red sign <sign_1> displaying a circular emblem <emblem_1> on the left, a line of Georgian script <Ka_Text_1> on top, and English text <En_Text_1> below. Along the ledge behind the sign, several small flags on thin poles <flags_1> are visible. To the left stands a portion of an older, traditional building <building_2> with a slender, tiered tower topped by a metallic spire and a five-pointed star.
Prompt reproduced word for word from the Black Forest Labs page named on this card.
What FLUX 3 Image is
FLUX 3 Image is Black Forest Labs' image model, released on 1 October 2026 as part of the FLUX 3 family. It generates from text, edits from a written instruction, and combines up to ten reference images in one request.
What separates it from the models beside it on this site is where you can put things. Any element you describe can carry a box, written as four numbers on a 0 to 1000 grid. The model builds the element inside that box and leaves the rest of the frame where it found it.
Star Task, the service this site calls, publishes the product as a preview rather than a finished release, so the service underneath can still change. This page separates what the published API accepts from what this workspace actually gives you a control for.
Where FLUX 3 Image sits in the FLUX 3 family
FLUX 3 is one model trained across image, video and audio together, announced on 23 July 2026. Black Forest Labs released it in stages, and the parts are documented separately.
The dates are worth keeping if you are comparing sources. FLUX 3 was announced on 23 July 2026, video arrived on 4 August 2026 as a preview, and image generation followed on 1 October 2026. A page that still describes the image model as unreleased is out of date.
Part
What it produces
How it is reached
FLUX 3 Image
Stills from text, instruction edits, and up to ten reference images
Hosted API, including this site
FLUX 3 Video
Video with synchronised audio
Hosted API. Not offered on this site
FLUX 3 Action
Actions for robots, simulators and games
Open weights, 7 billion parameters
What the published API accepts
FLUX 3 Image is priced and sized by a single resolution value, and the API accepts five of them. Black Forest Labs lists all five on its pricing page, while the developer documentation's own price table lists only four: 1.5k is missing there. Both figures are below, because anyone comparing the two sources will find the gap.
resolution
Output size
List price per image
Launch offer, 1 to 8 October 2026
768sq
768x768
$0.041
$0.0205
1k
About 1 megapixel
$0.048
$0.024
1.5k
About 2 megapixels
$0.07
$0.035
2k
About 4 megapixels
$0.10
$0.05
4k
About 16 megapixels
$0.607
$0.3035
The fifteen aspect ratios, and what auto follows
aspect_ratio takes fifteen fixed values from 21:9 to 9:21, or auto. auto follows the first image you send. With no reference images it produces a 1:1 frame, so a text-to-image request that should not be square needs the ratio set explicitly.
Ratio
Shape
Suits
21:9, 2:1
Ultra-wide
Cinematic landscapes, panoramas, website banners
16:9
Widescreen
Scenes with horizontal movement, web headers, slides, video thumbnails
3:2, 7:5, 4:3
Landscape photo
Street and documentary photography, interiors, groups of people
5:4, 1:1, 4:5
Near square
Products, single portraits, social posts, logos and icons
3:4, 5:7, 2:3
Portrait
Posters, magazine and book covers, full-length fashion
9:16, 1:2, 9:21
Tall
Phone screens, stories, vertical banners, towers and waterfalls
What this workspace exposes
The controls listed here are the ones this site actually renders. A parameter the upstream API publishes but the workspace has no control for is named as not exposed rather than quietly left out.
Two rows need a note. The resolution picker offers three tiers, not five. The workspace builds that control from a local list that admits 0.5K, 1K, 2K and 4K, and both 768sq and 1.5K fall outside it, so neither can be selected here even though the API accepts them. safety_tolerance and grounding are a different case: the product catalog publishes both, and this workspace has no control for either.
Control
In this workspace
Published upstream
Modes
Text to image, image to image, reference to image, image edit
The same four
Aspect ratios
Auto plus the fifteen fixed values
The same sixteen
Resolution
1K, 2K, 4K
Five tiers: 768sq, 1K, 1.5K, 2K, 4K
Reference images
Up to 10
Up to 10
safety_tolerance
Not exposed
0 to 4, default 2
grounding
Not exposed
On by default
Billing
Quoted before you submit, resolved by the service
Per image, by resolution tier
Bounding boxes: placement you can specify
A box is four integers in [top, left, bottom, right] order on a 0 to 1000 grid, so 500 is the middle of either axis. Every element you want to place gets one box, named in the prompt inside angle brackets, and the boxes are appended to the prompt you wrote.
One box per element is the documented rule, and the three roles are published with their own short descriptions: new generates at the target box, move carries an element from the source box to the target box, and anchor keeps it at the same position.
Describe the whole picture first and name each element you will place, as in "a black silhouette of a person <silhouette_1> against a solid chartreuse background <background_1>".
Give every named element one box, writing the four numbers in [top, left, bottom, right] order on the 0 to 1000 grid.
Decide what each box does. new generates at that box, move takes an element out of one box and puts it in another, and anchor holds an element exactly where it is.
Anchor everything that must not change. An anchored box carries the same numbers in both positions, and it needs its own description, because that description is what pins the element down.
How local an edit really is
Black Forest Labs measured how much of each edited frame came back identical to the source. It is the only published evidence for the claim that a boxed edit stays inside its box, and the nine measurements range from 67.8% to 94.1%.
Read the range honestly. Recolouring hair changed nearly a third of the frame, because hair is most of a portrait; adding two divers to a wide underwater scene changed almost nothing. A box narrows where the model works, and how much of the frame that accounts for still depends on the subject.
Published example
Edit type
Pixels identical to the source
Remove a cat
Remove
80.7%
Add a bird
Add
86.4%
Swap the plate
Replace
88.4%
Cat print on a T-shirt
Add
89.7%
Replace a rubber duck
Replace
85.5%
Recolour hair blue
Recolour
67.8%
Add a motorcyclist
Add
89.2%
Swap armchairs for a sofa
Replace
76.5%
Add two divers
Add
94.1%
Grounding and safety tolerance are published upstream but not reachable here
Two parameters in the published FLUX 3 Image interface have no control in this workspace. They are named here so that nobody assumes a switch exists because the API documents one.
Not having a control is not the same as the feature being absent. The service decides either way, and this page simply cannot change it.
grounding runs a web and image search before generating, which is what lets a prompt refer to something the model would not otherwise know. The published default is on, and turning it off is documented as faster and more literal.
safety_tolerance runs from 0, the strictest, to 4, the most permissive, and defaults to 2.
Neither can be set from this page. Requests made here are handled with whatever the service applies on its own.
There is no negative prompt
FLUX 3 Image has no negative_prompt field and no prompt upsampling field. The published guidance is to describe the scene you want instead of naming what you do not want, and Black Forest Labs supplies a replacement list for the phrases people reach for first.
Instead of
Write
no people
empty, deserted, solitary
no colors
monochrome, black and white, grayscale
no text
clean surfaces, unmarked, blank
no background clutter
plain studio backdrop in one color
no modern elements
traditional, historical, period-accurate
not dark
brightly lit, sun-drenched
not many
few, single, minimal
not running
walking slowly, standing still
Prompts you can copy
Each prompt below is published by Black Forest Labs in its own documentation and is reproduced word for word. Opening one loads it into the workspace.
The quickstart brief, a risograph rabbit
A two-color risograph narrative illustration with porous ink, soft misregistration, and cheerful simple shapes: a rabbit rides a pumpkin along a winding path through three layers of rolling hills. No text.
Two red pandas on bamboo
Two red pandas resting on bamboo beams, one looking relaxed and the other asleep. Their reddish-brown fur and white facial markings stand out against the soft, blurred natural background, suggesting a peaceful, sunny environment.
A mountain range at sunset
A breathtaking mountain range bathed in the warm glow of a sunset. The peaks, partially covered in snow, reflect hues of orange and pink. Fluffy clouds hover above, blending into a sky that transitions from golden to soft blue.
Prairie dogs, shallow depth of field
Three prairie dogs stand upright on a small dirt mound, facing the camera. The background is a blurred, sunlit field with golden hues, emphasizing the warm lighting. The prairie dogs appear alert, with a few green plants and scattered dry grass visible around them.
A waterfall from above
An aerial view of a waterfall cascading through a dense, verdant forest. The water flows down a rocky cliff into a pool, creating mist around the base. Lush, dark green trees surround the waterfall, contrasting with the lighter green grass visible in the bottom left.
Clownfish under warm light
Two vibrant clownfish swimming together in a dimly lit aquarium. The water reflects warm, earthy tones, and the background features soft orange coral formations, creating an underwater atmosphere.
FLUX 3 Image is not f.lux, Flux PCB or FLUX.2
Searching for this model's name on a major engine returns about as many unrelated results as relevant ones, so here is the list of near misses.
f.lux is a screen colour application for monitors. It sets the colour temperature of a display and has nothing to do with image generation.
Flux is a printed circuit board design tool, and Flux is also the name of a machine learning library for the Julia language.
FLUX.2 is the previous generation of this family. It is a different model on a different endpoint, and it is still served.
FLUX.1, together with its Kontext and Fill variants, is two generations back. The code repository that carries the Black Forest Labs name on GitHub belongs to that generation.
FLUX 3 Video is the sibling that produces video. This page covers still images only.
What this page does not promise
These are the limits of what is offered here, stated before anything is spent.
No free generations and no sign-up allowance. Every run is charged.
No open weights. FLUX 3 Image is reached through the hosted API only. The open-weights part of this family is FLUX 3 Action.
No file download from this page. Results are shown in the workspace.
No negative prompt field, because the model does not have one.
No control for grounding or safety tolerance, for the reason given above.
No 768sq or 1.5K output, because the resolution picker does not offer them.
No batched or queued generation. One run at a time.
No guarantee that a boxed edit leaves everything else untouched. The measured figures above show where it does not.
No claim that the prices on this page are what you will be charged. The workspace shows the cost before you submit, and that number is the one that binds.
Preview status. The service may change without notice, and this page documents a snapshot read on 3 October 2026.
Where the numbers come from
Every figure on this page has a source, and every source was read on 3 October 2026.
Source
What it supplies
Reference
Black Forest Labs documentation
The request surface, the bounding-box format, the aspect ratios, the resolution tiers, and the negative-prompt replacement list
The source photograph and the result after one cat was removed, published by Black Forest Labs as its worked example of a boxed removal. It measured 80.7% of the frame as identical to the original.
Before and after pair published by Black Forest Labs in its bounding-box documentation for FLUX 3 Image, read on 3 October 2026. · Official vendor sample reused at the site owner's request to document what the model does. No stock or authorship claim is made over it.
A layout example where the text itself is the subject. FLUX 3 Image renders lettering from the prompt rather than compositing it, which is why the same request can be reused across languages by changing what is quoted.
Sample published by Black Forest Labs in its bounding-box documentation for FLUX 3 Image, read on 3 October 2026. · Official vendor sample reused at the site owner's request to document what the model does. No stock or authorship claim is made over it.
A portrait-format layout example. Every object on the cover was named and boxed in the prompt, which is how a cover that has to place masthead, image and cover lines can be described rather than collaged.
Sample published by Black Forest Labs in its bounding-box documentation for FLUX 3 Image, read on 3 October 2026. · Official vendor sample reused at the site owner's request to document what the model does. No stock or authorship claim is made over it.
Questions
What is FLUX 3 Image?
It is Black Forest Labs' image model, released on 1 October 2026 as part of the FLUX 3 family. It generates images from text, edits them from a written instruction, and combines up to ten reference images in a single request.
When was it released, and is it finished?
FLUX 3 was announced on 23 July 2026, video arrived on 4 August 2026, and image generation followed on 1 October 2026. Star Task publishes this product with a preview availability flag rather than a general-availability one, so the upstream service can still change.
How much does one image cost here?
The workspace quotes the cost before you submit, and that quote is what binds, because the charge is resolved by the service rather than by this page. Star Task publishes a starting point of 6 points per run for this model, rising to 18 points for a 4K image.
What does Black Forest Labs charge for it?
Its own list prices are $0.041 at 768 x 768, $0.048 at 1K, $0.07 at 1.5K, $0.10 at 2K and $0.607 at 4K, with a 50% launch offer running from 1 to 8 October 2026. Those are the vendor's rates for its own API, not this site's prices, and this site charges in points.
Which resolutions can I pick here?
Three: 1K, 2K and 4K. The published API accepts five, and the two missing from this workspace's picker are 768sq and 1.5K.
Why are two resolution tiers missing?
The resolution control is built from a local list that admits 0.5K, 1K, 2K and 4K, and both 768sq and 1.5K fall outside it. Widening that list changes shared workspace code that every image model on this site uses, so this page documents the gap instead of pretending the control offers five tiers.
How many aspect ratios are there?
Sixteen on this page: auto plus fifteen fixed ratios from 21:9 to 9:21. auto follows the first reference image and falls back to 1:1 when there are none.
How many reference images can I attach?
Up to ten, which is the ceiling the product publishes for every mode it offers. Each image has to be at least 256 x 256 pixels, and anything above 4 megapixels is downscaled to 4 megapixels before it is used.
How do bounding boxes work?
You name each element inside angle brackets in the prompt and give it one box, written as four integers in [top, left, bottom, right] order on a 0 to 1000 grid. The boxes are appended to the prompt. Each one is marked new, move or anchor, and anything that must stay put should be anchored with its own description.
Can I use a negative prompt?
No. The model has no negative_prompt field and no prompt upsampling field. If something keeps appearing that you do not want, describe what should be there instead, which is what the replacement list above is for.
What is grounding?
It is a parameter, published as on by default, that runs a web and image search before generating so that a prompt can refer to something the model would not otherwise know about. The API publishes it and this workspace has no control for it, so it cannot be switched here.
Is this the same Flux as the screen colour app or the circuit board tool?
No. f.lux changes a monitor's colour temperature, and Flux is a printed circuit board design tool. Neither has any connection to FLUX 3 Image, which is a Black Forest Labs model. The name collides, and search results collide with it.
Explore details ↗ 
