State-of-the-art image generation and editing models, served via API, playground, and open weights
Flux is a family of AI image generation and editing models from Black Forest Labs for developers, creative teams, and enterprises.
AI Panel Score
6 AI reviews
Reviewed
AI Editor ApprovedApproved and published by our AI Editor-in-Chief after full panel analysis.Flux is a family of image generation and editing models from Black Forest Labs, available through a pay-as-you-go API, a browser playground, and downloadable open weights. It serves developers, creative teams, and enterprises that build image generation into production applications and creative workflows. Pricing is usage-based with no subscriptions or seat fees: hosted FLUX.2 models run from $0.014 per image for [klein] 4B to $0.07 for [max], and the 4B [klein] weights are free under Apache 2.0. Capabilities include multi-reference generation that keeps characters and styles consistent across hundreds of assets, mask-based object removal with FLUX Erase, image extension with FLUX Outpainting, and catalog-scale virtual try-on with FLUX VTO. The models also power generative editing in Adobe Photoshop and are available on Microsoft's Azure AI Foundry. Flux fits teams that want frontier image quality with self-hosting flexibility; alternatives include Midjourney, Stable Diffusion, Ideogram, and OpenAI's GPT Image.
Flux is used through three surfaces. The browser playground runs the most advanced FLUX models for testing ideas, iterating on prompts, and transforming images without code. The FLUX API exposes each model as its own endpoint (such as /v1/flux-2-pro) that takes a text prompt plus up to ten reference images and returns generated or edited output, billed per megapixel. Teams that need full control download the open-weight models from Hugging Face or GitHub and run them on their own infrastructure, with fine-tuning and LoRA customization rights available under commercial licenses.
The FLUX.2 family ships in five hosted variants: [klein] 4B and 9B for generation and editing in under a second, [flex] and [pro] in the middle of the range, and [max] as the flagship. FLUX.2 emphasizes production consistency, holding the same character and style across hundreds of generated assets, and product placement where lighting and physics adapt to the target scene. Dedicated tools extend the core models: FLUX Outpainting extends images beyond their frame in any direction, FLUX Erase removes masked objects along with their shadows and reflections, FLUX VTO renders virtual try-ons across thousands of catalog products, and FLUX Deblur sharpens blurry images. The earlier FLUX.1 Kontext line handles in-context editing that maintains character consistency across multiple editing turns.
Flux targets developers and product teams integrating image generation into applications, agencies producing creative assets at volume, and enterprises with data-sovereignty and compliance requirements (Black Forest Labs holds SOC 2 Type II and ISO 27001:2022 certifications). Pricing is purely usage-based with no subscriptions or seat fees: hosted generation runs from $0.014 per image for FLUX.2 [klein] 4B to $0.07 for [max], self-hosting licenses come in Builder, Platform, and Professional tiers, and the 4B [klein] weights are free under Apache 2.0. In the AI image generation category Flux competes with Midjourney, OpenAI's GPT Image, Stability AI's Stable Diffusion, Ideogram, and Google's Imagen.
Flux runs wherever the workload lives: the hosted API for production scale, on-premises deployment from open weights, Microsoft's Azure AI Foundry for enterprise procurement, and consumer hardware, with FLUX.2 [klein] shipping optimized for ASUS ProArt laptops in partnership with NVIDIA. An official MCP server at mcp.bfl.ai connects the models to AI agents, FLUX.1 Kontext powers generative editing inside Adobe Photoshop, and developer documentation lives at docs.bfl.ai.
Exposes each model as its own REST endpoint (e.g. /v1/flux-2-pro) with per-megapixel billing, built for production workloads at any scale.
Supports fine-tuning via the FLUX Pro Finetuning API, and open-weights licenses include fine-tuning and LoRA rights.
Model weights are downloadable from Hugging Face and GitHub for deployment on your own infrastructure; FLUX.2 [klein] 4B is licensed under Apache 2.0.
Removes masked objects together with their shadows and reflections and reconstructs the scene behind them.
Extends an image beyond its original frame in any direction while preserving lighting, texture, and composition.
Performs local edits and full scene transformations from text and image inputs, maintaining character consistency across multiple editing turns; powers generative editing in Adobe Photoshop.
Accepts up to 10 reference images per request to guide generation and edits, such as combining a product from one image with a scene or style from another.
Holds the same character and visual style across hundreds of generated assets for production sets.
Renders a person wearing a chosen garment with sub-4-second generations across thousands of catalog products.
Generates images from text prompts across five hosted FLUX.2 variants: [klein] 4B, [klein] 9B, [flex], [pro], and [max].
Places a product into a new scene with lighting and physics adapted to the target environment.
Runs FLUX models in the browser for prompt testing, iteration, and image transformation without code.
Free open weights for the 4B-parameter FLUX.2 [klein] model under the Apache 2.0 license, downloadable for self-hosting.
Cheapest hosted FLUX.2 variant; the first generated megapixel costs $0.014 and each subsequent megapixel $0.001.
Larger fast variant; the first generated megapixel costs $0.015 and each subsequent megapixel $0.002.
Professional-grade hosted model; the first generated megapixel costs $0.03 and each subsequent megapixel $0.015.
Flexible hosted model charged at $0.05 per megapixel on both reference images and the generated image.
Flagship hosted model; the first generated megapixel costs $0.07 and each subsequent megapixel $0.03.
Per-megapixel image tools billed alongside the core models: Outpainting, Erase, Virtual Try-On, and Deblur.
Self-hosting license for developers and early-stage teams, purchased through the BFL dashboard; list price not published.
Self-hosting license for product teams shipping at volume, purchased through the BFL dashboard; list price not published.
Self-hosting license for agencies and service providers, purchased through the BFL dashboard; list price not published.
Custom pricing with volume discounts, SLA guarantees, and dedicated support; requires contacting sales.
License for generating AI training datasets with FLUX.2 models; sold through the BFL dashboard.
The team that built Stable Diffusion now has a $300M war chest and an enterprise story.
“Black Forest Labs is a well-capitalized frontier lab with founding pedigree buyers can defend upward. The usage-based-only pricing is honest but hands you a variable line item instead of a fixed one.”
The founders shipped Stable Diffusion before this. That's most of the due diligence done — this isn't a first-time team learning generative imaging in public.
The capital backs it up. A $300M Series B in December 2025 at a $3.25B valuation, with NVIDIA, Salesforce Ventures, and a16z on the cap table, buys a defensible three-year runway. SOC 2 Type II and ISO 27001 mean procurement and legal won't stall the deal.
But billing is the real variable. Pricing is purely per-megapixel — FLUX.2 [max] at $0.07 a shot — so there's no seat cost to model and no ceiling to forecast either. Midjourney gives you a flat subscription; Flux gives you control and a variable bill. Pilot it on the hosted API, watch a month of usage, then decide on open-weights licensing.
Sits alongside Midjourney and OpenAI's GPT Image as a credible frontier option peers watch.
Founders behind Stable Diffusion plus SOC 2 and ISO 27001 make this an easy board defense.
The hosted API and browser playground let a team test before signing anything.
Adds a production image-generation capability that advances the product, not just a cost swap.
A $300M Series B at a $3.25B valuation with NVIDIA and a16z backing signals multi-year runway.
Product teams adding image generation who can defend a frontier-lab vendor to their board.
Buyers who need a fixed, predictable monthly cost with no metered billing.
Per-model endpoints and Apache 2.0 weights make Flux a bet you can walk back.
“Flux gives a platform team three deployment surfaces and an open-weights escape hatch most rivals don't offer. The self-hosting license model is powerful but genuinely complex to reason about.”
Each model ships as its own endpoint — /v1/flux-2-pro, /v1/flux-2-max — rather than one 'latest' route that silently changes underneath you. That's an architecture decision that respects version pinning, and it tells me the team thinks like infrastructure people, not a demo shop.
The strategic depth is the deployment optionality. You can call the hosted API today, then lift the same family to on-prem open weights when data-sovereignty rules tighten — FLUX.2 [klein] 4B is Apache 2.0, so the floor is genuinely free. Stability AI pioneered open weights here, but Black Forest Labs pairs them with a real commercial API and an MCP server at mcp.bfl.ai.
The tradeoff is licensing surface. Builder, Platform, and Professional self-host tiers each carry different image caps and domain limits, and none list a price publicly. If we standardize on Flux, someone owns tracking which weights we're licensed to run.
Sits with Stable Diffusion and Google Imagen at the frontier, differentiated by hosted-plus-open reach.
The API-plus-open-weights shape matches how senior ML teams actually ship image generation.
An MCP server, Photoshop via FLUX.1 Kontext, and Azure AI Foundry cover most stacks.
Apache 2.0 weights mean adopting Flux creates an exit path, not a lock-in.
Per-model endpoints and a size-distilled FLUX.2 [klein] show deep engineering, not a wrapper.
Platform teams who want hosted convenience now and a self-host exit path later.
Teams who want one simple license and no version management overhead.
Midjourney bills a monthly seat; Flux bills $0.014 to $0.07 per image, nothing idle.
“Hosted pricing is fully published and pay-as-you-go, which procurement will wave through. The compounding per-megapixel meter and unpublished self-host license prices are where the real cost hides.”
Midjourney charges $10 to $120 a month per seat. Flux charges nothing until you render. FLUX.2 [klein] 4B runs $0.014 for the first megapixel, $0.001 after. [max] tops out at $0.07.
Run the math. 100,000 [pro] images a month at roughly $0.03 each is about $3,000 monthly, $36K a year. Input images bill too — $0.015 per reference megapixel on [pro]. Editing-heavy pipelines with ten references per call add up faster than the sticker suggests.
No subscription, no seat tax, no auto-renewal trap — rare and welcome. But the self-host Builder and Platform licenses list no public price, so on-prem TCO needs a sales call. Compare against OpenAI's GPT Image before you commit; per-image economics differ by workload.
Hosted billing is frictionless, but unpublished self-host license prices reintroduce procurement drag.
Pay-as-you-go with no subscription, seat tax, or auto-renewal to negotiate out of.
Every hosted per-megapixel rate is published, no sales call to see the sticker.
Cost per image is measurable, though output-quality-to-price ROI varies by use case.
Per-megapixel plus per-reference-image billing compounds in editing-heavy pipelines.
Finance teams who prefer metered usage cost over fixed per-seat subscriptions.
Buyers who need a predictable fixed monthly invoice they can forecast exactly.
Swapping FLUX.2 [pro] for [max] means changing a URL, which cuts both ways.
“The endpoint-per-model design and up-to-ten reference inputs make Flux pleasant to script against. The per-megapixel meter means you build cost logging before you build features.”
Every model is its own REST endpoint, so calling /v1/flux-2-pro versus /v1/flux-2-max is a URL change, not a parameter flag. Clean for pinning a version in code; annoying when you want to A/B two models in one loop and now juggle two routes.
The reference-image handling is the daily win. Up to ten input images per request means product-in-scene composition is one call, not a chained pipeline. Character and style consistency holds across a batch, which is the part that usually breaks when you generate a hundred assets on Stable Diffusion locally.
Docs live at docs.bfl.ai and the MCP server drops Flux straight into an agent loop. The real friction is billing: every call charges per output megapixel and per input megapixel, so you write usage logging on day one. Replicate hides that behind flat per-run pricing, but Flux makes you watch every megapixel.
Endpoint-per-model and per-megapixel billing shape the daily loop once the demo glow fades.
Docs at docs.bfl.ai plus per-endpoint structure read like they were written for integrators.
The per-output and per-input megapixel meter demands cost logging from day one.
Fine-tuning via the FLUX Pro Finetuning API and LoRA rights give real advanced headroom.
Up to ten reference images per call folds composition into one request.
Engineers who script image generation and want per-model version control.
Developers who want flat per-run pricing without tracking megapixel usage.
FLUX Erase pulling shadows and reflections out with the object is the feature that sells it.
“The editing tools feel built by people who've actually fought messy real-world photos. It's web-and-API first, though, so this is a workshop for builders more than a lounge for browsers.”
FLUX Erase is the one that made me sit up. Most object-removal tools leave a ghost where the shadow was; this one pulls the object out with its shadow and reflection and rebuilds what's behind. FLUX Outpainting extends a frame in any direction and keeps the lighting believable.
The playground lets you poke at the real FLUX.2 models in a browser with no code, which is the right way to earn trust before you wire anything up. Character consistency across a hundred generated assets is the quiet workhorse — the thing that makes a campaign look like one shoot, not ten.
The catch: this is a builder's tool wearing a creative hat. There's a playground and an API, but no mobile app and no community gallery the way Midjourney has one. Ideogram feels friendlier if you just want to type and get a poster. Flux rewards people who'll push it.
Editing tools like FLUX Erase handle shadows and reflections most rivals leave behind.
The playground is approachable, but the real depth lives behind the API and licensing.
It's web-and-API first with no mobile app, which fits the builder audience.
The browser playground lets you test real FLUX.2 models before writing any code.
Sub-second [klein] generations and production-consistency claims suggest a solid, batch-ready core.
Creative builders who want deep editing control inside their own tools.
Casual users who want a mobile app and a browse-and-remix community.
Real founders, real certs, brutal category — Flux earns a watch, not a blank check.
“The pedigree and the compliance paperwork are the real deal. The worry is a category where Google and OpenAI ship monthly and pricing pressure never lets up.”
The founders built Stable Diffusion, then left Stability AI. Stability nearly came apart after they walked. So the talent is real. Whether lightning strikes the same lab twice is the open question.
Credit where it's due. The compliance paperwork is real — SOC 2 Type II, ISO 27001:2022 — and the $300M Series B in December 2025 buys runway. And the exit story is genuinely clean — FLUX.2 [klein] 4B ships under Apache 2.0, so if this goes sideways you keep running the weights. Most rivals can't say that.
The worry isn't the product. It's the neighborhood. Google's Imagen and OpenAI's GPT Image ship on a monthly cadence, and per-megapixel pricing in this space only goes one direction. FLUX.2 is barely seven months old. Strong hand — but a table where everyone's raising.
Hosted-plus-open-weights reach differentiates, but Midjourney, Imagen, and GPT Image crowd the space.
Apache 2.0 FLUX.2 [klein] weights let you keep running the model if the vendor shifts.
Strong funding and cadence, though a seven-month-old flagship in a monthly-shipping category is unproven.
The 'frontier' language is backed by shipped models and published compliance, not vapor.
Founders' Stable Diffusion pedigree fits winners, but the lab itself is only two years old.
Teams who want a frontier image model with a clean open-weights exit.
Buyers who need a long, proven track record before committing.
Common questions answered by our AI research team
Flux API pricing is pay-as-you-go with no subscriptions or seat fees. FLUX.2 [klein] 4B starts at $0.014 per image, [pro] costs $0.03, and the flagship [max] is $0.07 for the first generated megapixel.
Yes. FLUX.2 models unify generation and editing, FLUX Erase removes masked objects and rebuilds the scene behind them, and FLUX Outpainting extends images in any direction while preserving lighting and composition.
Yes. Open weights are downloadable from Hugging Face and GitHub. FLUX.2 [klein] 4B is free under Apache 2.0, and Builder, Platform, and Professional licenses cover commercial self-hosting with fine-tuning and LoRA rights.
Yes. Black Forest Labs holds SOC 2 Type II and ISO 27001:2022 certifications, and publishes usage, responsible-AI development, and training-data disclosure policies on its site.
Yes. FLUX.1 Kontext powers generative editing inside Adobe Photoshop. Flux models are also available on Microsoft's Azure AI Foundry, and an official MCP server at mcp.bfl.ai connects Flux to AI agents.
Company
Black Forest LabsFounded
2024Pricing
Usage-basedFree Plan
AvailableBlack Forest Labs is a Germany-based AI lab that develops FLUX, a text-to-image generation model, and other generative media models.
