FLUX.2 has three entrances and one rule that breaks most first attempts. Use the playground to watch it work, the API to build with it, or the open 4B weights to run it at home — and write every prompt as if negative prompts did not exist, because FLUX.2 does not support them.
What FLUX.2 is and where you can use it
Four tiers share the family name. The model page calls [max] top-tier quality for professional-grade output, [pro] state-of-the-art quality at maximum speed, [flex] maximum precision with text that reads, and [dev] open weights you deploy yourself. The overview adds [klein], pitched as fast and efficient for rapid iteration.
The published workflow claims are concrete: output up to 4MP, any aspect ratio, usable drafts from 400px², JSON-based control, pose guidance, 32K text input tokens, and sub-10-second generation. Multi-reference input takes up to 10 images in the playground and up to 8 through the API.
Three ways to make your first image
Playground first. playground.bfl.ai needs no key and covers [max], [pro] and [flex]. The official wording is explicit about what is missing: [klein] is not there, and is reachable through the API and Hugging Face instead.
The API route is four steps — create an account at dashboard.bfl.ai and verify the email, add credits, create a key, make the first call. Credits sell at 1 credit = $0.01 USD, and the key is shown once. Generation is asynchronous: you POST a request, get back a polling_url, then poll it until the status reads Ready and the image sits at result.sample. The API walkthrough covers that loop end to end.
Local is the third door, and it only opens for the Base weights: the overview states they ship as open weights for local development and research, and are not offered on the public API.
The prompt skeleton: four blocks
The prompt guide offers one framework and asks you to reuse it: Subject + Action + Style + Context. Subject is the main focus, action is what it does, style is the artistic approach or medium, and context is setting, lighting, time and mood.
Order carries weight. FLUX.2 pays more attention to what comes first, and the guide’s priority runs main subject, key action, critical style, essential context, then secondary details. Length scales with ambition: 10–30 words for quick concepts, 30–80 for most projects, 80 or more for complex scenes.
For photorealism, name equipment. The guide’s examples include Shot on Hasselblad X2D, 80mm lens, f/2.8, natural lighting, and film stocks work the same way: shot on Kodak Portra 400, natural grain, organic colors.
A baker in a flour-dusted apron pulling a tray of sourdough from a stone oven, steam rising, warm tungsten light from the left, documentary photography, shot on Kodak Portra 400, natural grain, quiet early-morning moodAspect ratios and resolution limits
| Ratio | Published sizes |
|---|---|
| 1:1 Square | 1024×1024, 1536×1536 |
| 16:9 Widescreen | 1920×1080, 1536×864 |
| 9:16 Portrait | 1080×1920, 864×1536 |
| 4:3 Classic | 1536×1152, 1024×768 |
| 21:9 Ultrawide | 2048×864 |
The hard rules sit underneath that table: a minimum of 64×64, a maximum of 4MP, dimensions in multiples of 16, and a recommendation to stay at or below 2MP for most work. The quick reference gives [flex] a guidance of 4.5 within a 1.5–10 range and a step ceiling of 50.
prompt_upsampling rewrites a thin prompt for you, and it is worth knowing where it stops: the overview states that [klein] does not include it. Short prompts and [klein] are a bad pair.
Two worked examples: product and poster
Example one: the product shot
A 4:3 canvas at 1536×1152, no reference image, default guidance. The prompt carries subject and context, then hands rendering to one camera line.
A cream knit sweater folded on a pale oak table beside a ceramic cup, soft window light from the right, minimalist e-commerce product photography, shot on a 100mm macro lens, f/8, seamless light-grey background, high dynamic rangeExample two: the poster
A 9:16 canvas at 1080×1920, because posters are read upright. Text is the hard part: wrap the words in quotation marks, say where they sit, name the typeface feel, and attach colors as hex codes beside the word color, since a vague use #FF0000 somewhere produces inconsistent results.
A minimalist travel poster for an overnight train through the Alps, the headline text "NIGHT LINE" in large bold sans-serif letters across the top, the text "Zürich to Milano" in small body copy along the bottom edge, deep navy background with a single warm horizon glow in color #FF6B35, flat vector illustration, generous empty spaceWhen reference images are involved, describe each one’s job — subject from image one, style from image two, background from image three. The [pro] API shares a 9MP ceiling across input and output, so a 2MP result leaves room for seven. If a product or character must survive every generation, a fine-tuned klein adapter beats reference images.
Three prompt mistakes worth fixing
The first is describing what you do not want. The prompt guide opens with No negative prompts and asks you to spend those words on what should be there instead.
The second is writing three words and shipping it on [klein]. That variant has no upsampling step behind it, so detail has to come from you.
The third is assuming the official guide has a Chinese example to copy. Its multi-language section documents French, Thai and Korean prompts and carries no Chinese one, so Chinese prompts work but the vendor publishes no quality claim for them. The fix is the same in any language: be specific across all four blocks. The prompt-structure guide goes deeper.
Weak: a cat, cute, 4k, no blur, no watermark
Better: a grey tabby cat asleep on a windowsill, soft morning light through lace curtains, shot on Kodak Portra 400, natural grain, calm domestic moodFrequently asked questions
Does FLUX.2 accept negative prompts? No. The prompt guide says so in its opening line.
Can I prompt in Chinese? Yes, the model takes it, and the official multi-language section documents French, Thai and Korean examples but no Chinese one, so the vendor publishes no quality claim.
Where should a beginner start? The playground, because it needs no key and covers [max], [pro] and [flex]. [klein] is API and Hugging Face only.
What does one image cost? The overview lists [klein] from $0.014 per image, [pro] from $0.03 per MP and [max] from $0.07 per MP; for [flex] the two official pages disagree at $0.06 and $0.05 per MP, and both figures are published. The free-tier and pricing breakdown covers what is not.