Flux in Batch Banana: best uses and how to prompt it

A magazine cover titled VERSO: a model with a black bob in a red wool coat whose collar fans out in stiff pointed petals, on a backdrop graded from yellow to violet, with three small cover lines at the lower left

The masthead and all three cover lines came out as written. The backdrop and the hard light came from a reference image we can't show here, which is why the prompt mentions Image 1.

Best for

Flux is Black Forest Labs' family of image models. Start on Flux 3 Image. It makes an image from a prompt or changes the ones you attach, sets lettering in many languages, and can look up real places and products on the web before it draws. It reads up to 10 reference images, and they add nothing to the price.

Reach for it when the photo and the type come out together, as on a magazine cover or a tea tin, for a real landmark drawn as it is, and for your own product in a new scene.

For long exact copy and dense layouts, try GPT Image. For a file you can scale and edit, use Arrow or Recraft. Which model for what matches models to jobs.

Which version

Model	Best for	Credits per image	Reference images
Flux 3 Image	Generation and editing in one model, with web search	6	Up to 10
Flux 2 Pro	Top-tier Flux quality with multi-reference support	7	Up to 8
Flux 2 Max	Exceptional realism, precision, and consistency	17	Up to 8
Flux 2 Klein 9B	Enhanced realism, crisper text and better color adherence	4	Up to 5

Start on Flux 3 Image, at 6 credits for a 1K image. Use Flux 2 Klein 9B to find a composition fast. It costs 4 credits and takes about 4 seconds on average. Flux 2 Max is the highest-quality Flux 2, built for photoreal scenes with natural skin and light. Flux 2 Pro costs less than half as much as Max.

Pro and Max add detail of their own to a short prompt. Klein draws only what you write, so give it full sentences and the light. Black Forest Labs' advice for Klein: "describe scenes like a novelist".

The same prompt on each version:

A woman in a sand trench coat strides across a wet Paris crosswalk, carrying a brown handbag, passers-by blurred on either side

A woman in a sand trench coat and matching trousers crosses a Paris street, the long coat flaring in the wind, cars and lights soft behind

Only Flux 3 Image blurred the passers-by, as the prompt asked. Max and Pro softened the street behind her instead, and Klein drew the crowd sharp.

The model menu also has Flux 2, Flux 2 Flex, Flux 2 Klein 4B and Flux Kontext Max. Models lists them with their prices, along with Flux 3 Video, the family's video model.

Write the prompt

Start short, then add what you want to control. Flux 3 Image expands a short prompt into a detailed description before it draws. "A fisherman mending a net" came back as a whole harbour, with boats, ropes and a red jacket the prompt never mentioned. Add detail only where you care, and put first whatever matters most. The second prompt names the crop, the net's color, the tool and the light, and gets them:

An old fisherman in a red jacket sits on a crate on a busy harbour quay, mending a green net among ropes, buckets and moored boats

Close-up of an old fisherman's hands pulling a wooden netting needle through a bright orange net in low dawn sun, a blue boat hull soft behind

Name the camera and the light, not the praise. Write "Shot on Fujifilm X-T5, 35mm f/1.4", not "professional photo", and give the light a source and a direction. "Highly detailed" and "ultrarealistic" add little.

Quote the words, place them, describe the type. Put every word in quotation marks, with the capitals and punctuation you want, then give its place, weight, case and color. For another script, keep the prompt in English and quote the script directly:

A square metal tea tin on a sunlit wooden shelf, its lid printed with the kanji 朝霧 in black brush strokes above MORNING MIST · SENCHA in small capitals

Both lines came out exactly as quoted, the kanji and the small English line under it.

Say what's there, not what isn't. Flux 3 Image has no negative prompt, so write "an empty promenade", not "no crowds".

Tie each hex code to one thing. "The lid is #2F5D50" colors the lid. A bare "use #2F5D50" lets the model choose.

Web search

With Web search on, Flux 3 Image searches the web and images for what your prompt names before it generates. Name a real landmark, city or product, and it works from what it finds instead of from memory.

Web search is on by default, and only Flux 3 Image has it. You'll find it in the model menu, under Flux 3 Image's details.

The model menu with Flux 3 Image highlighted. Below its details, the Web search switch is on.

Name the real thing plainly, "the Flatiron Building at dawn", and let the search fill in how it looks. With Web search off, the same prompt drew a wide stone building that isn't the Flatiron:

The narrow wedge of the Flatiron Building at dawn under a pink sky, a coffee cart with a striped awning on the wet corner in front of it

A wide, squat stone building at dawn that is not the Flatiron, a coffee cart with a striped awning steaming on the wet pavement in front of it

Turn it off when everything in the prompt is made up, a hare in a waxed coat included, and you'll get results faster.

Edit and combine images

Attach an image and your prompt becomes an instruction for it. There is no edit mode to switch on. Flux 3 Image reads up to 10 images, each at least 256 pixels on a side, and they cost nothing extra.

Describe the change, not the whole scene. "Make the wrapper cloudberry orange" is the whole prompt. Then name what must stay, as Black Forest Labs does in their own example: "Keep the puddle, grass, butterfly, low camera viewpoint and warm golden light unchanged."

A Halvard chocolate bar in a cobalt wrapper printed with fine white contour lines and HALVARD in a tall cream serif, lying across an old atlas of the Norwegian coast in hard window light, its last squares of dark chocolate showing

The Halvard bar on the atlas with its wrapper turned a muted coral, the contour lines, the HALVARD lettering, the bare squares and the light unchanged

Everything on the keep list held: the contour lines, the lettering, the bare squares, the atlas and the light. The color did not match: cloudberry orange came back a muted coral. A color name leaves the exact shade to the model, so give a hex code when the shade matters.

Give each image a role, and call it by its number. "Image 1" is the first image you attached, "image 2" the second.

Fern the hare in an olive coat sits on a snowy bench by a fjord, breaking a square off a cobalt Halvard bar, drawn in ink and watercolour

Fern is image 1 and the Halvard bar is image 2. The bar came through with its cobalt wrapper, contour lines and name, redrawn in Fern's ink and watercolor.

Settings

Resolution on Flux 3 Image has four steps: 1K (the default), 1.5K, 2K and 4K. Black Forest Labs suggests 1K for most web and app images, and 2K for "print, hero images, and crops that need detail". 4K costs 72 credits, twelve times 1K, and Black Forest Labs says it "can take several minutes".

Find the image at 1K, then run the prompt again at 2K for the one you keep. A new resolution is a new image, not an enlargement, so expect a different take on the same prompt.

Every Flux version takes all ten aspect ratios, from 21:9 to 9:16.

Model	Credits per image	Other settings	Each reference image adds
Flux 3 Image	6	Resolution: 1.5K 9, 2K 12, 4K 72	0
Flux 2 Pro	7	Resolution: 2 MP 11, 4 MP 18	9
Flux 2 Max	17	Resolution: 2 MP 25, 4 MP 39	17
Flux 2 Klein 9B	4	Resolution: 0.25 MP 1, 0.5 MP 2, 2 MP 8, 4 MP 16	2

The Flux 2 versions charge for each reference image you attach, shown in the last column. Flux 3 Image doesn't.

Sources

Black Forest Labs' Flux 3 guide: basics, building a prompt, style and text, resolution, editing and the quick reference, which covers Flux 2.