What it is
Flux is a family of AI image models made by Black Forest Labs, a young company based in Freiburg im Breisgau, a city on the edge of the Black Forest in south-west Germany. You describe a picture in words, such as "a lighthouse on a cliff at dusk, painted in watercolour", and a Flux model produces it in a few seconds. Newer versions can also edit a picture you give them, following an instruction like "change the jacket to red" or "put this character in a snowy street".
Black Forest Labs announced itself and its first models, FLUX.1, on 1 August 2024. Its best-known founders, Robin Rombach, Andreas Blattmann and Patrick Esser, are researchers who earlier helped create the technology behind Stable Diffusion, one of the first image generators anyone could download. That background shows in how Flux is released: some versions are sold as a paid online service, and others can be downloaded and run on your own computer.
Flux is not usually an app you open in the way you open Midjourney. It is more like an engine that other products build on. You may have used it without knowing: it has appeared inside chatbots, design apps and photo editors, and anyone can reach it through Black Forest Labs' own website, its API (a way for one program to use another), or free demo pages.
How it works, in plain terms. Like most image generators, Flux starts from random visual noise, a bit like television static, and turns it step by step into a picture that matches the words. That is the idea explained on the Diffusion Model page. The FLUX.1 model cards describe the models as "rectified flow transformers": a refinement of the same idea that learns a more direct path from noise to finished image, so fewer steps are needed, combined with the transformer design that also powers chatbots. The original FLUX.1 models each have 12 billion parameters, the learned numbers that make up a model.
The versions. Black Forest Labs names each model with a label in square brackets, and the labels tell you how you are allowed to use it. At a high level, using the company's own announcements:
- [pro], [max] and [flex]: the company's most capable versions, which it keeps to itself. You pay per image through its website, its API or partner services.
- [dev]: downloadable "open weights", free to use for non-commercial purposes under Black Forest Labs' own licence, with paid licences for business use.
- [schnell] and [klein]: smaller, faster downloadable versions ("schnell" is German for fast, "klein" for small) released under the Apache 2.0 licence, a standard open-source licence that allows commercial use.
The generations so far.
- August 2024: FLUX.1 in three versions: [pro] through the API, [dev] under a non-commercial licence, and [schnell] under Apache 2.0.
- October 2024: FLUX1.1 [pro], which the company said was six times faster than the first [pro], and the launch of its own paid API.
- November 2024: FLUX.1 Tools, add-ons for filling in or extending part of a picture and for following the outline or depth of a reference image.
- May 2025: FLUX.1 Kontext, for editing images with plain-language instructions, plus a web "playground" for trying the models. A downloadable [dev] version followed on 26 June 2025.
- November 2025: FLUX.2 (25 November), in [pro], [flex] and [dev] versions. The [dev] model is much larger, at 32 billion parameters, and the company said it can combine up to ten reference images and handle complex lettering. A [max] version followed in December.
- January 2026: FLUX.2 [klein] (15 January), small models in 4-billion and 9-billion parameter sizes, designed to make an image in under a second on a good graphics card.
- July 2026: FLUX 3 (announced 23 July), which moves beyond still pictures to video with sound, at first in limited early access.
This line-up will change, probably within months. Check Black Forest Labs' announcements and release notes for the current list.
Why it matters
It raised the bar for downloadable image models. When FLUX.1 arrived in August 2024, reviewers and users widely remarked on its realistic photos, its ability to follow long prompts, and its ability to write readable words inside pictures, which older generators often turned into gibberish. Some of those strengths were also available in versions anyone could download. For people who want to run image AI on their own computer, for privacy, cost or control, Flux became one of the main choices.
It is the engine inside other people's products. Black Forest Labs makes most of its money by licensing its models to other companies. Over two years Flux has turned up in X's Grok chatbot, Mistral's Le Chat, Adobe Photoshop and Meta's apps, among others. So even if you never visit Black Forest Labs' website, understanding Flux helps you understand what is happening inside tools you already use, and why their terms and quality change when a company swaps one engine for another.
It shows that "open" comes in shades. The Flux family is a clear example of the difference between open weights and open source. Open weights means the company lets you download the finished model files. Open source, in the strict sense, means a licence that lets anyone use, change and share the software for any purpose, including business. Only some Flux models meet that stricter bar. If you plan to use AI images in anything you sell, this difference matters, and Flux is a good place to learn it.
It gives creators more control over the same pictures. Editing tools like Kontext let you keep a character, product or setting consistent across several images by showing the model a reference picture, something early image generators struggled with. For an author planning a series of covers, a small shop photographing products, or a creator making a comic, consistency is often the difference between a toy and a useful tool.
It raises the same hard questions as every image generator. Realistic pictures can be used to deceive. Black Forest Labs, like its rivals, has not published a full list of the images its models learned from; its transparency page points instead to the summaries that Californian and European law require. And the legal status of AI images, both whether they infringe on others' work and whether they can be protected themselves, is still being worked out in courts and copyright offices.
Real-world examples
Launch week: a new lab, and pictures on X
On 1 August 2024, Black Forest Labs came out of stealth, announcing 31 million dollars of seed funding led by the investment firm Andreessen Horowitz alongside its three FLUX.1 models. Two weeks later, on 13 August, Elon Musk's xAI released Grok-2, and the image generator inside Grok, available to paying X users, turned out to be powered by Flux. VentureBeat reported on 14 August that Grok's pictures were far less restricted than those from Midjourney or OpenAI's DALLĀ·E 3, and that users had quickly made compromising images of public figures, including the US presidential candidates Kamala Harris and Donald Trump. The partnership did not last: in December 2024 xAI switched Grok to its own image model, Aurora. The episode is worth remembering for two reasons. It showed how quickly a small lab's model could reach millions of people through someone else's app, and it showed that the safeguards a user sees are often set by the app, not only by the model maker. The Grok page tells that story in more detail.
Kontext: editing with words, and a place in Photoshop
On 29 May 2025, Black Forest Labs released FLUX.1 Kontext, a set of models that take a picture plus an instruction and change only what you ask. Its announcement highlighted keeping a character recognisable across new scenes, editing text on signs, and making small local changes without redrawing the whole image. The company claimed Kontext was much faster than rival editing models; as with any company's own benchmark, treat such claims as the maker's view. From 25 September 2025, Adobe offered FLUX.1 Kontext [pro] as one of the outside "partner models" in the beta version of Photoshop's Generative Fill, alongside Google's Gemini image model and Adobe's own Adobe Firefly models. For a Photoshop user, that meant choosing between engines from inside the same tool. It also marked a shift for Adobe, which had until then built its generative features on its own Firefly models.
From three researchers to a billion-dollar supplier
Black Forest Labs grew quickly by selling to other companies rather than to the public. In November 2024, Mistral's Le Chat assistant added image generation using Flux Pro. On 9 September 2025, Bloomberg reported that Meta had agreed a multi-year deal worth 140 million dollars to use Black Forest Labs' image technology. Later that month, when Meta launched Vibes, a feed of short AI-generated videos in its Meta AI app, Meta's chief AI officer Alexandr Wang wrote that the early version used Midjourney and Black Forest Labs while Meta built its own models. On 1 December 2025, TechCrunch reported that the company had raised 300 million dollars at a valuation of 3.25 billion dollars, with investors including Salesforce Ventures, NVIDIA, and the design companies Canva and Figma's venture arm. The lesson for readers is practical: when a big company's image feature suddenly changes style or quality, the reason is often a new engine supplier behind the scenes, and those supply deals can end.
An everyday case: cover concepts for a self-published novel
This is an illustrative scenario, not a real person. Priya is about to self-publish a cosy fantasy novel and has a small budget for a professional cover designer. Before she books one, she uses a free Flux demo to explore ideas: a lantern-lit bookshop in the rain, a fox curled on a stack of books, a teal-and-gold colour scheme. Twenty minutes and thirty images later she has a clear direction and a folder of mood pictures. She also notices the problems: the fox has five legs in one image, and the shop sign spells her title slightly wrong. She sends the best three images to her designer as a brief, and the designer creates the final cover, with proper typography and a licence Priya understands. If she had instead used a Flux image as the cover itself, she would need to check the licence of the exact model she used and, when publishing on Amazon's Kindle Direct Publishing, to disclose that the cover was AI-generated, which its rules require at the time of writing.
Try it yourself
You need a web browser and about 20 minutes. These steps use Black Forest Labs' free public demo of FLUX.2 [klein] 4B, hosted on Hugging Face, a site that hosts AI models. At the time of writing it ran without an account, but free demos can be busy, limited for visitors who are not signed in, or taken down. If it does not work, Black Forest Labs' own playground at playground.bfl.ai offers its top models, but needs an account and uses credits, with some free credits for new users at the time of writing. Check the current terms either way.
- Open the demo. Go to huggingface.co/spaces/black-forest-labs/FLUX.2-klein-4B and wait for it to load. You will see a Prompt box, a Run button and an area for the result.
- Start with a simple prompt. Copy and paste this, then press Run:
A cosy bookshop at night in the rain, warm light in the windows, a black cat sitting on the step, photograph - Look closely. Check the fine details: the cat's paws, the lettering on any sign, whether the reflections in the puddles make sense. Flux is known for realistic results, but every image model still makes small mistakes, and spotting them is the most useful skill you can build.
- Test its lettering. Ask for words in the picture, one of the things Flux is known for:
A vintage book cover design with the title "THE LANTERN KEEPER" in gold letters, a lighthouse in a storm, painted illustrationCheck the spelling letter by letter. Short titles usually work better than long ones. - Change one thing at a time. Keep the same prompt and change a single word, for example "photograph" to "watercolour painting" or "night" to "dawn". Comparing results tells you what each word is doing. This is the same iterative refinement you would use with a chatbot.
- Fix the seed to compare fairly. Open Advanced Settings, untick "Randomize seed" and type a number such as 42. The seed decides the starting noise, so with the same seed, differences come from your words, not from luck.
- Try an edit. If the demo shows an "Input image(s)" section, upload a photo you took yourself, such as a mug on your desk, and write an instruction like
Make this mug bright yellow and add steam rising from it. Only use photos you own and never photos of other people without their permission. - Check the licence before using anything. The klein 4B model's weights are released under the Apache 2.0 licence, which allows commercial use. Other Flux models have different terms, and any website that runs Flux for you has its own terms of service as well. Read both before you sell or publish.
- Optional: run it on your own computer. If you have a recent graphics card, you can download open Flux models and run them offline with free software such as ComfyUI. The FLUX.2 [klein] 4B model card says it fits in about 13 GB of graphics memory, on cards such as the NVIDIA RTX 3090 or 4070 and above, and the larger models need more. The Parameters page explains why bigger models need bigger hardware.
Common mistakes and misunderstandings
Thinking "open" means "free for any use". Only some Flux models use a true open-source licence ([schnell] and FLUX.2 [klein] 4B, under Apache 2.0). The [dev] models, and the larger klein 9B, are open weights under Black Forest Labs' non-commercial licence. You can download them, but the licence says running them for "revenue-generating activity" or for business use is not allowed without buying a licence from the company.
Misreading what the non-commercial licence says about pictures. This one confuses many people. The FLUX.1 [dev] licence, at the time of writing, says Black Forest Labs claims no ownership of the images you generate and that you "may use Output for any purpose (including for commercial purposes)", except that you may not use them to train a competing model. But the same licence limits how you may run the model itself: personal research, hobby and private use are fine, while use connected to a business or earning money is not. In practice, a hobbyist's image is on firmer ground than one made as part of a paid job or a product. If you plan to earn money, use a model with a commercial licence, a paid service whose terms cover you, or buy a licence. Licences change, so read the current text for the exact model you use, and ask a lawyer if a lot is at stake.
Assuming an AI image is automatically yours to protect. Being allowed to use an image is not the same as owning its copyright. In the United States, the Copyright Office's January 2025 report concluded that prompts alone do not give a person enough control over the result to count as its author. A federal appeals court upheld the rule that copyright needs a human author in March 2025, and the Supreme Court declined to hear an appeal in March 2026. Your own changes, drawing and arrangement of AI material can still be protected. Rules differ by country, so check where you publish.
Forgetting disclosure rules. Some platforms require you to say when you have used AI. At the time of writing, Amazon's Kindle Direct Publishing asks authors to disclose AI-generated images, including covers, even if they edited them afterwards. Check each platform's current rules before you upload.
Copying a living artist's style or a real person's face. Asking for a picture "in the style of" a named living illustrator, or of a real person, raises legal and ethical problems, and Black Forest Labs' licences forbid uses such as non-consensual intimate images and infringing content. Describe the look you want in your own words instead: the colours, the brushwork, the mood.
Expecting perfect realism. Flux is known for convincing photographs, but it still produces extra fingers, impossible reflections, misspelled lettering and objects that melt into each other. Look at every image at full size before you use it.
Assuming the same name means the same model. "Flux" covers many models of very different sizes and abilities. A small, fast model on a free website will not match the company's top paid model, and a website that says "powered by Flux" may be using an older version. When comparing, find out which version you are actually using.
Believing the maker's benchmarks. Black Forest Labs, like every AI company, publishes comparisons in which its own models look good. Independent leaderboards move constantly, and taste in images is personal. Try several tools, such as Flux, Midjourney, OpenAI Images and Ideogram, on your own prompts before deciding.
Related elements
- Stable Diffusion: the earlier downloadable image model that Flux's founders helped create.
- Diffusion Model: the noise-to-picture technique that Flux refines.
- Midjourney: a popular subscription image generator, a useful comparison for style and ease of use.
- OpenAI Images: OpenAI's image generation inside ChatGPT, another common alternative.
- Ideogram: an image generator also known for putting readable words in pictures.
- Adobe Firefly: Adobe's own image models, which sit beside Flux Kontext in Photoshop.
- Grok: xAI's chatbot, which used Flux for its first image generator in 2024.
- Parameters: why model size decides what hardware you need to run Flux at home.
- Fine-tuning: how people teach open Flux models a particular style, product or character.
- Guardrails: the content filters that the Flux licences require and that apps set on top.