Work with me
image Black Forest Labs

Flux 2 Pro

Black Forest Labs' production image model takes a structured prompt instead of a paragraph, matches an exact hex colour, and charges a published rate per megapixel.

At a glance

Resolution
64x64 up to 4 megapixels, every edge a multiple of 16
Aspect ratios
presets from 1:1 to 21:9, plus any custom size
References
up to 8 through the API, inside a shared 9 megapixel budget
Inputs
text or JSON, plus reference images

Good for

  • Jobs where a program writes the prompt, not a person
  • Brand work with a fixed palette, because you can give a hex code
  • Product composites built from several reference photos
  • Costing a job before you run it, at a published rate per megapixel

Weak at

  • Using all eight references at a large output size, because of the shared budget
  • Output above 4 megapixels, where newer models go higher
  • Quality or sampling dials, which the pro variant does not expose
  • Being the current model, since its maker has moved on to a newer family

What it is for

Flux 2 Pro is a still image model built for repeat work. It generates and it edits, in the same model, and it is happiest when the same kind of job runs again and again.

Two things make it different from the other image models on this site. You can send the prompt as structured JSON instead of prose. And you can name a colour as a hex code instead of describing it. Both matter when a machine is filling in the prompt, or when a brand palette is fixed and “deep navy” is not good enough.

The third thing is the price. Black Forest Labs publishes a rate per megapixel, so you can work out the cost of an image before you make it. Most of its rivals charge by token, which you can only estimate.

This page is written from published documentation. I have not run work through Flux 2 Pro myself, so nothing here is a judgement on how the pictures look.

What it takes in

Text, or JSON, plus reference images.

Output runs from 64 by 64 pixels up to 4 megapixels, which is about 2048 by 2048. Every edge has to divide by 16. The docs suggest staying at 2 megapixels or below for most work.

There are shape presets from square up to a 21:9 ultrawide, and you can also ask for a custom width and height instead.

The reference limit is 8 through the API. You will see 10 repeated in a lot of places. That figure is the playground, and it also applies to a different variant. The API gives you eight slots.

How to prompt it

Write the prompt as a form, not a story. Scene, subject, style, colours, lighting, composition, camera. The model reads that structure well in plain text, and it reads the same thing as JSON if you would rather send JSON.

For colours, put the hex code next to the object it belongs to. The docs are clear that a loose instruction like “use this red somewhere” gives loose results. For text in the image, write the words inside quote marks and say where they sit.

A product card with a fixed brand colour 
Scene: a single matte ceramic mug on a pale concrete surface, shot from
slightly above.
Subject: the mug is glazed in color #0F3D3E. A thin band near the rim is
color #E8C468.
Text: the word “RITUAL” printed small on the side of the mug, facing camera,
in a light geometric sans.
Lighting: one soft window light from the left, gentle shadow to the right.
Composition: centred, square crop, generous empty space around the mug.
Do not add extra props, extra text or a logo.

For an edit, attach the reference and describe the change in the same shape. Name what must stay first. There is no mask, so the “what stays” list is the only lock you have.

If you are pasting from a page like this one, the angle brackets are escaped here for the web. Type them normally in the tool.

You can run it on ImagineArt.

Where it falls down

The shared megapixel budget is the trap. Eight references sounds generous until you ask for a large image, and then the model has far less room for them. If you need many references, keep the output small and upscale later.

Four megapixels is the ceiling. That is fine for social and for most web work, and it is short of what GPT Image 2.5 or Nano Banana Pro will hand you for print.

The pro variant hides its sampling controls. There are no quality levels to turn up when a result is close. The flex variant exposes guidance and steps if you want those dials, but that is a different model to switch to.

It is also no longer the newest thing its maker sells. Flux 2 shipped in November 2025 and Black Forest Labs has released newer work since. It is still supported and still on four of our five platforms. Expect improvements to land elsewhere.

A few numbers are simply not published. There is no confirmed prompt length limit, and the docs never list which image file types you can send in. Test both on your own account rather than trusting a figure you read somewhere.

Vosu names Flux without saying which version, so its pricing page is not proof you get this one.

Sources