FLUX is Black Forest Labs’ image model line, and the generation its pages document today is FLUX.2 — the family behind FLUX.2 [pro], FLUX.2 [flex] and FLUX.2 [klein]. It is a family rather than one checkpoint, so “how to use FLUX AI” is two questions: which tier, and where you run it.
This guide covers the tiers, the routes, a first image step by step, one worked prompt with references and hex colours, and the figures the vendor does not publish.
What FLUX.2 is
The model page’s headline is FLUX.2 [pro] | [flex], and its description line covers the generation: production-grade image generation and editing, 4MP photorealistic output, multi-reference control.
Five tiers are documented, and each trades quality, speed, control and delivery against the others.
| Tier | What Black Forest Labs documents it as |
|---|---|
| FLUX.2 [max] | Top-tier generation and editing, with grounded generation from web search |
| FLUX.2 [pro] | “State-of-the-art quality at maximum speed. Both, not either.” — the production tier |
| FLUX.2 [flex] | Maximum precision, with adjustable inference steps and guidance scale |
| FLUX.2 [klein] | The fastest family, documented at sub-second inference |
| FLUX.2 [dev] | Open weights, documented as local-only with no hosted API |
The release post describes one architecture behind them: a Mistral-3 24B vision-language model coupled with a rectified flow transformer, editing images at up to 4 megapixels.
One generational note: Black Forest Labs announced FLUX 3 in July 2026 as a multimodal model covering image, video and audio. FLUX.2 is still the image family its documentation covers.
Where you can use FLUX.2 today
Four routes are documented, and they are not interchangeable.
| Route | What the official pages say about it |
|---|---|
| Playground | A browser demo at playground.bfl.ai, documented as needing no signup or card |
| Official API | Endpoints called with a key created in the dashboard |
| MCP server | The official server at mcp.bfl.ai, connected over OAuth instead of a key |
| Self-hosted | The open-weight [dev] tier, on your own hardware |
An account starts in the dashboard: register at dashboard.bfl.ai, confirm the email, add credits, create a key. The official warning is worth repeating — the key is shown once and must never sit in client-side code.
The FLUX.2 API getting-started guide continues with the request itself, and the FLUX.2 topic hub collects the family’s remaining documentation. Prices stay out on purpose: sites that resell FLUX.2 on their own credit plans set their own, and those are not Black Forest Labs’ prices.
Your first image, step by step
The first run proves the loop works; it is not the run for your best image.
- Pick the tier first. [pro] for production, [flex] when lettering decides, [klein] while exploring.
- Order the sentence the official way: subject, action, style, context. FLUX.2 pays more attention to what comes first.
- Size the frame on purpose: sides in multiples of 16, a 4MP ceiling, 2MP suggested.
- Bind every colour to an object. A bare hex code does little; the guide says it works best when clearly associated with a specific object.
- Change one clause per round and keep the rest identical, so the round shows which word moved the image.
There is no negative prompt to write, because the guide states that FLUX.2 does not support them. Every “no X” becomes a description of what should fill the frame. The habit comes from older open-weight families, and the Stable Diffusion guide is where most writers picked it up.
Worked example: one product shot and one reference-built scene
Both cases use the same four-part sentence; the tier is the only variable.
Case one needs lettering to survive: a glass jar of capsules, and a label that must match the reference exactly.
Subject: a glass jar filled with capsules standing on a matte concrete plinth, centred, the label facing the camera and reading "BFL" in a cream serif
Action: the jar stays still; nothing moves in frame
Style: studio product photography, one large soft key light from the upper left, a single soft rim light on the right, seamless cool grey background, a soft contact shadow under the jar
Context: a reference image of the label is supplied; keep the letterforms, the cream colour and the small pine motif on every capsule identical to that reference
Colour: the label cream is #F2E8D5 and the pine green is #1F4D3A; apply each only to the object it belongs to
Size: 2048x2048, 4MP, no text anywhere except the labelWhat to inspect afterwards is narrow: the letterforms, the motif on every capsule, and whether the grey background picked up warmth from the reference. Small type usually fails first, which matches the limitation on the model card — text rendered may be inaccurate or subject to distortion. Moving the label sentence to the front does more than adding description, because word order is the lever the guide names.
Case two is the scene in the figure above. Three references, three plates, one light:
Subject: a snow leopard, a red panda and a penguin, each sitting on its own stacked television
Action: all three hold still and look toward the camera
Style: documentary photography, one key light from the front left covering all three plates, cool grey backdrop, matching grain across the frame
Context: image one is the snow leopard, image two is the red panda, image three is the penguin; keep each animal's markings and colouring identical to its own reference, and keep the lighting shared rather than per-plate
Colour: the backdrop grey is #B9BDC1
Size: 2048x1152The failure mode there is the light rather than the animals: each reference arrives with its own, so the prompt has to say the light is shared. That is what multi-reference control is for, and where the tier shows — the overview documentation lists up to 8 references through the API and up to 10 in the playground, while [klein] is documented at up to 4. If typesetting is most of your work, the GPT Image guide covers the other line we document.
A prompt structure that holds up
The two cases share one skeleton, which is what makes a result repeatable instead of lucky.
The blocks worth writing every time
The official framework is four parts long, and the practical blocks below sit under it. Length guidance from the same guide: 10–30 words, 30–80, or 80 and up.
Subject: who or what is in frame, written first
Action: what changes, or that nothing changes
Style: medium, light sources and their direction, surface and grain
Context: what each reference image is for, and what must stay identical
Colour: each hex code attached to the object it belongs to
Size: the dimensions, kept to multiples of 16 and at or under 4MPNegative clauses have no home in that list, because the model does not read them. Rewrite the absence as a presence: name the background that fills the frame.
Where FLUX.2 stops short
Black Forest Labs publishes its own limitations, which is rare enough to quote. Rendered text may be inaccurate or distorted. The model may fail to generate output matching the prompt. Prompt following is heavily influenced by the prompting style. [klein] adds one note: no prompt upsampling.
Limits, cost and what Black Forest Labs has not published
The hard limits are published: 4MP maximum output, a 64-pixel minimum side, dimensions in multiples of 16, and up to 2MP recommended. References are counted rather than free — each is charged as 1MP whatever the output resolution. The [pro] API budgets 9MP across input and output, so a 1MP output leaves room for 8 references and a 2MP output for 7.
Cost is credit-based and published: 1 credit is $0.01 USD, the API and the playground share one price, and [pro] starts at $0.03 per megapixel for text-to-image; batches multiply that by the image count.
What is missing matters as much. No free API quota is documented — the free demo the [klein] page points at is a playground trial, which is a different thing. No benchmark score or head-to-head comparison appears on the vendor’s pages. No unified parameter count is published for the hosted tiers.
Frequently asked questions
Do I have to pay to try FLUX.2? The playground demo is documented as free, with no signup and no credit card. A free API quota is not published.
Which tier should I start with? [pro] for production, [flex] when typography decides, [klein] when speed matters, [dev] when you run it yourself.
Can I use negative prompts? No — the guide states that FLUX.2 does not support them, so describe what should be in frame.
How many reference images can I send? Up to 8 through the API and up to 10 in the playground; [klein] is documented at up to 4.
Are the weights open? The tier documented as open weights is [dev], described on official pages as local-only with no hosted API. Weight licensing for the other tiers is documented per tier on the vendor’s own pages.
Is this out of date now that FLUX 3 exists? No. FLUX 3 is a newer multimodal generation covering video and audio too, while FLUX.2 remains the image family on the model page.
Learn the loop once — one clause per round, the shape named before the style, references sized against the budget — and FLUX.2 stops feeling like a slot machine.