Black Forest Labs Users Manual

As of …

A practical guide to FLUX 2 Pro — Black Forest Labs' frontier image-generation model. 32B-parameter Rectified Flow Transformer paired with a Mistral-3 24B Vision Language Model, generating and editing images up to 4 megapixels. The lab is run by veterans of the original Stable Diffusion team.

🎈 ELI5

Black Forest Labs (BFL) is a German AI lab founded by people who worked on the original Stable Diffusion. Their FLUX family is one of the most respected image-generation lineages outside the Big-3 clouds — strong on photorealism, prompt fidelity, and (increasingly) text rendering inside images.

FLUX 2 Pro is the current pro tier — 32B parameters, a vision-language model attached, up to 4MP output, and accepts up to 10 reference images per request for consistency.

Getting started in 60 seconds

Pick your door: bfl.ai for the official API, fal.ai, OpenRouter, attap.ai, and other inference partners host hosted variants.
Pick the model: flux.2-pro for the workhorse, flux.2-max for the highest-fidelity tier (where available).
Reference your way to consistency. FLUX 2 Pro accepts up to 10 reference images per call — the strongest way to get character / brand / scene consistency across a series.
Plan around 4MP. Native max is 4 megapixels; for larger outputs, upscale separately.

Which surface should I use?

BFL official API — cheapest at the per-image rate ($0.014/image base); best for production.
fal.ai — pay-per-megapixel; convenient for iterating in a UI.
OpenRouter — multi-vendor routing.
attap.ai — credit-priced (Flux 2 Pro at 2 credits).
Self-host — selected FLUX models have open weights for non-commercial / research use; FLUX 2 Pro is the commercial frontier tier with stricter licensing — verify before deploying.

FLUX 2 Pro — deep dive

Area	What FLUX 2 Pro does
Architecture	32-billion-parameter Rectified Flow Transformer integrated with Mistral-3 24B Vision Language Model for prompt understanding.
Resolution	Up to 4 megapixels per generation. Maintains accurate text rendering, color matching, and character identity at full res.
Reference images	Up to 10 reference images per generation request — the highest in this comparison.
Text in image	Approximately 60% accuracy on first attempt for complex typography. Iterate to refine; first-pass text is meaningfully better than older FLUX tiers.
Editing	Natural-language editing — describe the change, the model modifies the existing image in-place rather than starting from scratch.
Best for	Brand-consistent imagery, photorealism, layout-heavy creative, product hero shots, character series.

FLUX family

Tier	Use it for
FLUX 2 Pro workhorse	Default production tier. Best $/quality for most workflows.
FLUX 2 Max	Top-tier quality — photorealistic product shots and creative direction. Where available, the choice for hero assets.
FLUX 1.1 / 1.x lineage	Predecessor family; still hosted on third-party providers. Migrate to 2.x when convenient.

Release timeline

Date	Release	What changed
2024	BFL founded; FLUX.1 family	Founded by ex-Stable Diffusion team. Initial open-weight tier + closed Pro/Ultra tiers.
2024-2025	FLUX 1.1 series	Iterative improvements; better text and prompt fidelity.
Late 2025	FLUX 2 Pro	Frontier release: 32B + Mistral-3 VLM, 4MP, 10-image references, natural-language editing.
2026	FLUX 2 Max + ecosystem	Higher-quality tier; broader provider coverage (fal.ai, OpenRouter, attap.ai, others).

Pricing

Provider	Pricing	Notes
BFL official API	From `$0.014/image`	Cheapest at the per-image rate; production-grade reliability.
fal.ai	~`$0.03/MP` + `$0.015/extra MP`	Pay-per-megapixel; convenient if you batch many small images.
attap.ai	2 credits per generation	Bundled with other models; good for trying alongside competitors.
OpenRouter / others	Varies	Multiple hosts; pricing drift is common.

Account & keys

Visit bfl.ai; sign up.
Add a payment method; buy credits.
Generate an API key. Env vars only.

First API call

BFL's API is REST; submit a generation request, poll for completion, fetch the image URL. The fal.ai-hosted variant uses fal's standard async pattern.

curl — BFL API curl -X POST https://api.bfl.ai/v1/flux-2-pro \ -H "x-key: $BFL_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "prompt": "Editorial portrait, soft north light, 35mm", "width": 1024, "height": 1024 }' # Returns a polling URL; GET it until status == "Ready"

Editing & references

Two patterns make FLUX 2 Pro genuinely powerful for production:

Multi-image references — submit up to 10 images alongside the prompt to lock subject, style, color, layout. Strongest way to get a consistent "look."
Natural-language editing — submit the previous output as a reference + a description of the change. The model modifies in-place; a series of edits maintains coherence better than a series of fresh generations.

Reference economy Each reference image you submit costs latency and slightly more per-call compute. Use as many as you actually need — most series stabilize around 3-5 references rather than the full 10.

Prompt library

Photorealistic product hero

Product hero Editorial product photograph: [product] on [surface] with [lighting]. Style: [photorealistic / studio / outdoor]. Camera: [35mm / 50mm / 85mm], shallow depth of field. Lighting: [soft north light / golden hour / dramatic side-light]. Background: [out-of-focus / textured / negative space]. Output: 4MP, 16:9. References uploaded: [product reference, style reference, color palette reference].

Typography poster

Typography Movie-poster style design with the headline: "[exact text]" Layout: - Headline dominates the upper third. - Subhead "[exact text]" centered, smaller. - Credits block at bottom in tiny condensed type. Style: [reference: e.g. Saul Bass / minimalist / 1970s sci-fi]. Mood: [adjective]. Critical: render the text exactly. If you can't fit it cleanly, prioritize legibility over decoration.

Character series with consistency

Character series Generate the next image in this character series. Character references uploaded: [3-5 prior images] Style references uploaded: [1-2 mood references] This frame: - Action: [describe] - Setting: [describe] - Camera: [describe angle and framing] - Lighting: [describe] Match the character's appearance EXACTLY across face, hair, clothing, distinguishing features. Setting can change; character cannot.

Patterns

"Lock with references, vary with prompts"

Use the 10-image reference slots for the things that must stay constant (character, brand, color palette). Use the prompt for what should change (action, setting, camera). This split produces the most predictable series.

"Iterate via edit, not regenerate"

Once an output is close, refine via natural-language editing rather than re-prompting from scratch. Edits preserve micro-details a regeneration would scramble.

"Typography is a one-attempt-three-tries problem"

FLUX 2 Pro's ~60% first-attempt accuracy on complex typography means you'll often need 2-3 tries. Plan accordingly — don't anchor a deadline to the first generation.