Getting Started with FLUX

The image model that finally gets text right — but you'll reach it through someone else's front door.

Last reviewed 2026-10-14 · for FLUX

How it works

FLUX is a model — the AI "brain" that does the actual work — not a website you log into. Whichever route you reach it through, the motion is the same: describe the picture you want, look at what comes back, change the wording, go again.

What it is

An image and editing model family from Black Forest Labs, the German lab founded by the researchers behind Stable Diffusion. The current line is FLUX.2: [max], [pro], [flex], [dev] and [klein]. Two are open-weight — the model file itself is published, so you can run it yourself, which no closed rival allows. The rest sit behind a paid API, a way for programs to talk to each other without a person clicking.

Its pitch against Midjourney: legible text inside images, and tight instruction-following. BFL says typography, infographics and UI mockups "now work reliably", and that it holds a character consistent across up to ten reference images.

Before you start

Going direct to Black Forest Labs' dashboard at bfl.ai is pay-as-you-go, no subscription. You pay per image, at a few cents each, and per-model rates vary by resolution; the calculator on bfl.ai/pricing shows the current rate for the model and size you pick.

Most beginners should come in through someone else's app. Krea and Freepik are ordinary image-making websites — sign up, type, get pictures — that license FLUX and run it for you. Both have a free tier, you pay them rather than BFL, and their terms set your rights.

Running the files yourself is free, and the barrier is hardware, not skill: the freely licensed FLUX.2 [klein] 4B wants a graphics card with roughly 13GB of memory on it — most laptops don't have that — plus software like ComfyUI to drive it. Below that bar, a third-party app is the same model for cents an image.

Security & privacy

Privacy differs sharply by route. [klein] on your own machine is genuinely private — no server is involved. Every other route sends what you type to someone's computer, BFL's or a third-party app's, which may pass it on. Open source is not the same as private.

Free to download is not free to sell from

Images made through BFL's paid API are yours: their Developer Terms say you own the Output and may use it commercially.
FLUX.2 [klein] 4B is Apache 2.0 — free, and free to sell from. The 9B is under the FLUX Non-Commercial License: free to download, not for commercial use.
FLUX.2 [dev], the large open-weight model, is non-commercial by default; selling needs a paid self-hosted licence.
For FLUX these permissions genuinely come apart — check your variant before anything reaches a client invoice.

What it's good (and not good) at

Best at

Readable text inside an image — signage, packaging, poster headlines
Long, multi-part instructions, without quietly dropping half
Editing by instruction: change the jacket, leave the rest alone

Okay at

Painterly art — Midjourney is prettier unasked
Photorealistic faces: good, occasionally uncanny
Fast iteration on the small variants, trading fidelity for speed

Avoid

Expecting a polished consumer app with a gallery — BFL doesn't make one
Body text: headlines hold, paragraphs degrade into plausible nonsense
Anything commercial on a non-commercially-licensed variant — a legal limit, not a quality one

Starter prompts to try

Test the text-rendering claim first

A weathered enamel shop sign on a brick wall, cream lettering on deep green, reading exactly: HARBOUR ROAD BAKERY - est. 1974. Soft overcast daylight.

Why this one: Legible text is the capability BFL leads with for FLUX.2, so a failure here tells you about your route immediately.

Stack five requirements and count what survives

A ceramicist's workbench seen from directly above. Exactly three unfired bowls in the upper left. A blue cloth over the right edge. Warm light from the left. No hands.

Why this one: FLUX.2 is tuned for adherence to complex structured instructions, so counting how many hold is a fair test, not a gotcha.

Edit a picture with words instead of tools

Keep this photograph as it is, but change the parked car from silver to matte dark green. Change nothing else — not the background, the lighting, the reflections.

Why this one: FLUX.2 generates and edits in one model at up to 4MP, and change one thing, touch nothing else is the job it was built for.

Common pitfalls

It is not a website you log into. There is no consumer FLUX app the way there is a Midjourney — BFL runs a developer dashboard with a playground, billed against prepaid credit. A free browser tool with a gallery means a third-party product with FLUX inside.
Confusing the generations. FLUX.1, FLUX.2 and FLUX 3 are different things; FLUX 3 is a combined image, video and audio model, not a successor to the image line. Check the release notes when what you read doesn't match what you see.

Where to next

Image tools you can just sign up for Running open models yourself

Want the full picture first? See the FLUX overview for pricing, the free tier and related tools.