gianhossle.com/research/black-forest-labs
← all research
// research · company deep dive

Black Forest Labs: Founding Story and Product

How the researchers behind latent diffusion and Stable Diffusion built FLUX, and how Black Forest Labs turns frontier image and video models into a business.

Author
Gian Hossle
Date
September 2026
Origin
Self-initiated, out of my own interest
Subject
Black Forest Labs, Freiburg (Germany)
Format
12 pages · 16 figures · 2 tables
Sources
Linked inline: arXiv, bfl.ai, Bloomberg, Crunchbase News, Sacra
⟲ drag to rotate · scroll to zoom
Made in the Black Forest, built to run inside other people's products.Freiburg sits on the edge of the Black Forest, the home of the cuckoo clock: a precision mechanism people trust for the time it keeps, not for the name on the movement. FLUX plays the same role. It runs inside Photoshop, Le Chat, Canva and Picsart, promises control, answers in under half a second with FLUX.2 [klein], and has shipped three generations in two years.
Figure 1

From a university lab to FLUX 3

Research, product and business milestones, 2020 to 2026. Pick a dot or play the story.

202120222023202420252026the research yearsAug 2024 · out of stealth in Freiburg
Research6 milestones
Product8 milestones
Business6 milestones
Aug 2024product

Launch with FLUX.1

7 / 20

Sources: arXiv, bfl.ai, Wikipedia, TechCrunch, Bloomberg

// product marketing view

Two things Black Forest Labs got right

01

They put FLUX where the users already are

Growth runs through other companies' products and a free-to-start funnel, so it never depends on the Black Forest Labs name.

  1. Embedded, not a destinationMany people who use FLUX never see the Black Forest Labs name: they meet it inside Adobe Photoshop, Mistral's Le Chat, Canva or Picsart.p. 5
  2. Credibility on launch weekWithin two weeks of the August 2024 launch, FLUX powered image generation in xAI's Grok, and independent rankings put its quality ahead of Midjourney and OpenAI.p. 2
  3. Three doors, one model familyA pay-as-you-go API with no subscriptions or seat fees and a free Playground, open weights with paid commercial licenses, and enterprise options such as a managed API with zero data retention.p. 9-10
  4. Open weights as the funnelResearch and personal use is free, and FLUX.1 [dev] alone has about 15,000 likes on Hugging Face. Companies that run the weights inside their own products buy a license.p. 10
02

They sold control and finished outcomes, not just a model

Frontier quality, fine control and freedom over where the model runs: three things that are rarely available together.

  1. Control is the standout featureUp to ten reference images keep a character, product or style consistent across a campaign without training a custom model. Colors take exact hex codes and typography stays readable.p. 6
  2. A ladder, not one modelFLUX.2 comes in five variants that trade quality against speed and openness. The small [klein] makes an image in under half a second and became the first FLUX model to ship on consumer hardware, on ASUS ProArt laptops.p. 6-7
  3. Jobs packaged as one API callFLUX Tools turn the raw model into finished outcomes, such as Virtual Try On in under four seconds, that product teams can plug in without any prompt engineering.p. 9
  4. Numbers customers can repeatOne luxury brand went from 1,000 to more than 100,000 images per season, from two to four hours per asset to under a minute. Burda cut creation time by 50%.p. 11-12

The catch the paper names: a model that powers someone else's product can be swapped out, as Grok showed when it replaced FLUX with xAI's own Aurora model in December 2024.p. 11