Who Is Who: Black Forest Labs, the German Lab Inside Photoshop and Canva

Snapshot (as of 26 July 2026)

• Founded: August 2024, Freiburg im Breisgau, Germany, with a second base in San Francisco
• Employees: the company calls itself a 100 person team, company trackers listed around 107 in May 2026
• Main product: the FLUX family of image, video and action models
• Total funding: more than 450 million dollars
• Last valuation: 3.25 billion dollars, December 2025
• Main investors: Andreessen Horowitz, AMP, Salesforce Ventures, Nvidia, General Catalyst, Adobe Ventures, Figma Ventures, Canva, Deutsche Telekom’s T.Capital
• Revenue: around 96 million dollars annualised as of August 2025, according to an estimate by the research firm Sacra. The company publishes no figures
• CEO: Robin Rombach
• Available in the EU: yes, own interface plus Microsoft Azure, with computing capacity in Frankfurt

---

If you have used generative fill in Adobe Photoshop, made something in Canva, or asked Mistral’s chat assistant for a picture, there is a fair chance a model from a company in Freiburg im Breisgau did the work. Not Silicon Valley. A university town of 230,000 people at the edge of the Black Forest, closer to Basel than to Berlin.

Black Forest Labs turns two this month. It has the headcount of a mid sized dental practice chain and a valuation of 3.25 billion dollars. That makes it Germany’s most valuable company in generative media, though no longer the country’s most valuable AI company overall: the Munich defence software firm Helsing raised 1.8 billion dollars at an 18 billion dollar valuation on 13 July 2026.

Still, almost nobody outside the field can name Black Forest Labs. That gap between influence and recognition is the interesting part, and it is deliberate. This is a company that sells to the companies you have heard of.

Where it came from

The founding story starts with a departure.

In March 2024, several of the core researchers behind Stable Diffusion left Stability AI, the British company that had made image generation a mass phenomenon and was then, by wide reporting, in financial trouble. Among them were Robin Rombach, Andreas Blattmann, Patrick Esser and Dominik Lorenz. Their shared work goes back further, to research at LMU Munich on latent diffusion, the technique that made high quality image generation cheap enough to run on ordinary graphics cards.

That last point explains everything that follows. Latent diffusion did not just make better pictures. It made the whole thing affordable. Before it, generating a decent image meant industrial computing. After it, a gaming PC could do it.

Five months later, in August 2024, the team announced Black Forest Labs with 31 million dollars of seed funding, the first outside money a company takes, led by Andreessen Horowitz. The angel list gives away the network: Garry Tan of Y Combinator, Brendan Iribe of Oculus, the Hollywood agent Michael Ovitz, and Timo Aila from Nvidia’s research organisation.

The location was the statement. The team could have raised the same money in San Francisco with fewer questions. They kept the headquarters in Germany and opened a second office in California.

What they actually sell

Three things, in descending order of visibility.

Open weight models. Some versions of FLUX are published so anyone can download and run them. The naming is confusing and worth decoding: [dev] versions have open weights, meaning the trained model file itself is public, but the licence restricts commercial use. [schnell], German for fast, is the genuinely permissive one. [pro] and [max] stay closed. The FLUX.2 family runs from a 32 billion parameter open model down to [klein] versions of 4 and 9 billion parameters that need as little as 13 gigabytes of graphics memory, which puts them within reach of a well equipped home computer.

Pay per picture access. According to the company’s price list in July 2026, the small [klein] models start around 0.014 dollars per image, roughly 0.013 euros, and the premium tiers from about 0.03 dollars per megapixel of output. On Microsoft’s Azure marketplace, the older FLUX.1 generation is listed at 40 dollars per 1,000 images, while FLUX.2 is billed by megapixel. For a small business making a few hundred images a month this is coffee money. For a platform making millions, it is a real contract.

Enterprise licences and platform deals. This is where the money is, and it is invisible to consumers. Adobe Photoshop, Canva, Picsart, Burda, Krea and Deutsche Telekom all work with FLUX models. Reporting in September 2025, carried into Sacra’s modelling, put a multi year Meta contract at about 140 million dollars, with combined contract value across the large platform partners around 300 million. Those are outside estimates, not company figures.

On 23 July 2026 the company announced FLUX 3, a model trained on images, video and audio in one architecture and extended to predicting physical actions. It ships as FLUX 3 Video, Image, Action and Dev. Alongside it came FLUX-mimic, built with the Swiss robotics company mimic and, according to both companies, already being tested in production at Audi. The claim worth watching: a manipulation task that used to need 30 or more hours of robot demonstration data can be fine tuned in as little as 30 minutes.

The money

The funding history has an unusual feature. One of the rounds was kept quiet.

• Seed, August 2024: 31 million dollars, led by Andreessen Horowitz, with General Catalyst and Match VC
• Series A: the second institutional round, led by Andreessen Horowitz, with BroadLight Capital, Creandum, Earlybird, General Catalyst, Northzone and Nvidia. It was never publicly announced at the time and was disclosed only later. The company has not stated the amount, and the published totals imply something in the region of 120 million dollars
• Series B, 1 December 2025: 300 million dollars at a 3.25 billion dollar valuation after the new money was counted in, co led by Salesforce Ventures and AMP, with Andreessen Horowitz, Nvidia and General Catalyst joining

Total raised is above 450 million dollars.

On revenue, be careful. Sacra estimates roughly 96 million dollars in annualised revenue, meaning a recent month’s income multiplied out to a full year, as of August 2025. That is an outside estimate built from contract reporting and modelling, not an audited number, and the company publishes nothing. If it is roughly right, a company of about a hundred people was running at close to a million dollars of revenue per employee, which is the kind of ratio that explains a 3.25 billion dollar valuation. If it is wrong, the valuation is a bet on the models rather than the business. Both readings are available and neither can be settled from outside.

The cost side is easier. Training image and video models is compute intensive, video far more so than images. The company has computing capacity in Frankfurt and Virginia. Nvidia is both supplier and shareholder, an arrangement worth noticing whenever you see it in this industry.

Who owns it, who runs it

Nobody outside the company sees the share register, but the shape is legible, and it contains a detail most coverage misses.

Several of the biggest customers are also shareholders. Adobe Ventures, Figma Ventures, Canva and Deutsche Telekom’s investment arm T.Capital all appear in the company’s own investor list. So when this profile describes Adobe and Canva as platform partners, that is only half of it. They are partners with an equity interest in the supplier they depend on.

That cuts both ways. It makes those relationships stickier than a normal contract, which softens the customer concentration risk. It also means a handful of large platforms have a seat at the table when the company decides what to build and what to keep open.

American capital still sets the pace. Andreessen Horowitz led the seed and the Series A and joined the Series B. Salesforce Ventures and AMP co led the Series B. European investors are present, Creandum from Sweden, Earlybird from Germany, Northzone from the Nordics, T.Capital from Bonn, but none of them led a round.

One name runs through the whole story. Anjney Midha led the seed round as a general partner at Andreessen Horowitz. He left the firm in late 2025 to start his own fund, AMP, and then co led the Series B from there. The same person has backed this company across three rounds and two employers.

The founders keep operating control. Robin Rombach is CEO. That matters more than usual in a research led company, where the asset that could walk out of the door is the same asset that walked out of Stability AI in 2024.

The people who matter

Robin Rombach, co founder and CEO. First author of the latent diffusion paper that made modern image generation practical, later a research lead at Stability AI. He is the reason investors treat this team as the real lineage of Stable Diffusion rather than a copy of it. His public framing of the company is unusually specific: vision as the highest bandwidth signal about the physical world, and one architecture trained across images, video, audio and actions together.

Patrick Esser, co founder. Co author with Rombach of the earlier VQGAN work on high resolution image synthesis, which the company itself names as the start of the lineage. Later at Stability AI.

Andreas Blattmann, co founder. Worked on extending latent diffusion from still images into video during the Munich and Stability years, which is exactly the direction the company has now taken with FLUX 3.

Dominik Lorenz, co founder. The fourth member of the original research group, part of the same body of work that the open image generation ecosystem is built on.

That is the unusual thing here. This is not a business team that hired researchers. The founders are the researchers, and their models have been downloaded more than half a billion times according to the company.

Anjney Midha, investor and board level backer. Backed the company from a16z and then from his own fund AMP. Also sits close to Mistral. If you want to understand how a small number of people shape which European AI companies get to scale, he is a good place to start.

Strong where, weak where

Strong. The dual licence model works: give away enough that developers build on you, sell the rest. Open versions on Hugging Face create adoption, awareness and free improvement work, while the closed tiers and platform contracts collect the money. It is the Red Hat playbook applied to model weights, and Black Forest Labs runs it better than anyone else in images. The capital efficiency is genuinely rare: about a hundred people, a possible nine figure revenue rate, and models sitting inside products with hundreds of millions of users.

Weak. Three things.

Image generation is commoditising fast. When OpenAI, Google and Adobe bundle image generation into subscriptions people already pay for, standalone generation becomes a feature rather than a product, and price pressure follows.

The customer base is concentrated, and partners do leave. xAI used FLUX inside Grok, then moved to its own image model. Sifted reported in April 2025 that the collaboration had ended, and noted it was unclear exactly when.

And the video and robotics ambition is expensive. Moving from images to video with sound, and then to action prediction, multiplies computing costs at exactly the moment margins in image generation are under pressure. FLUX 3 is a bet that the company can climb into a higher value category before the lower one gets squeezed.

The European angle

This is where the profile differs from an American one.

Availability and processing. Models are reachable through the company’s own interface and through Microsoft Azure, with computing capacity in Frankfurt. That is a starting point for a European buyer, but it is capacity, not a promise. The company does not publish a contractual commitment to process only inside the EU, which is what a compliance officer would actually ask for.

Open weights change the compliance picture. Under the EU AI Act, providers of general purpose AI models carry documentation duties, but models released under a genuinely free and open licence are exempt from some of them, while still having to follow copyright rules and publish a summary of the content used for training. Black Forest Labs maintains a public training data disclosure page. Whether that satisfies every reading of the law is a lawyer’s question. The direction of travel is unusually transparent for this industry.

The inheritance nobody can fully price. The founders’ earlier work sits inside the copyright disputes of the first image generation wave, including cases brought against Stability AI in the United States and the United Kingdom. Those are cases against Stability, not against Black Forest Labs, and the distinction matters. But anyone buying image models in Europe is buying into an unsettled legal area, and that transparency page reads like a company that knows it.

And it is a German supplier to European industry. Deutsche Telekom is both customer and investor. Burda is testing FLUX 3. Audi is testing the robotics model. That is a narrower position than winning a consumer chat app, and a more defensible one.

What to watch next

Three things specific enough to check.

Does the open weight version of FLUX 3 actually ship this year? The announcement promised faster and open weight versions later in 2026. If that slips or arrives cut down, the balance is tipping from community towards enterprise, which would change what this company is.

Does the Audi test become a contract? Robotics is the reason for the FLUX 3 architecture. A named production deployment with volumes attached would confirm it. Continued silence after a launch announcement would suggest the opposite.

Where does the next round come from? Every round so far has been led from the US. A European led growth round would be a real signal about whether Europe can finance its own champions. Another American led round tells you the answer is still no.

Sources

• Black Forest Labs: FLUX 3 announcement (23 July 2026, company press release)
• Black Forest Labs: Laying the foundations for visual intelligence, our 300 million dollar Series B (1 December 2025)
• Black Forest Labs: Training data disclosure
• Black Forest Labs: Pricing (prices checked July 2026)
• Sacra: Black Forest Labs revenue, valuation and funding (revenue and contract values are Sacra estimates)
• Sifted: Elon Musk’s xAI no longer working with German startup Black Forest Labs (April 2025)
• Tech.eu: Helsing secures 1.8 billion dollar Series E at 18 billion dollar valuation (13 July 2026)
• Forbes: Key Stable Diffusion researchers leave Stability AI (20 March 2024)