Black Forest Labs Inc., a startup that develops artificial intelligence models optimized for image generation tasks, has raised $300 million in funding.

Salesforce Ventures and prominent venture capitalist Anjney Midha co-led the Series B round. They were joined by Nvidia Corp., Canva Pty, Figma Ventures, Temasek and over a half dozen venture capital firms. Black Forest Labs disclosed in its announcement of the deal today that it now has a post-money valuation of $2.25 billion.

The Germany-based company was founded last year by machine learning researchers Robin Rombach, Patrick Esser and Andreas Blattmann. The trio previously helped develop the Stable Diffusion series of open-source image generation models. Black Forest Labs offers its own open-source image generators that it says provide higher output quality than many alternatives with lower inference costs.

The company debuted its newest model series, FLUX.2, last month. The algorithms are not based on the diffusion architecture that underpins Stable Diffusion and many other image generators. Instead, it uses a newer design known as a flow matching architecture.

Flow matching models share several technical similarities with their diffusion-based counterparts. Both start the image generation process with a file that contains noise and gradually fill in details. The difference is that a flow matching model performs the task in a more hardware-efficient manner, which speeds up inference.

The flow matching model that powers FLUX.2 incorporates elements of the transformer architecture. That’s a neural network design most commonly found in large language models. Furthermore, Flux.2 includes an open-source vision model from Mistral AI SAS. According to Black Forest Labs, the latter algorithm makes the image generator series better at tasks such as rendering lighting effects.

Flux.2 doesn’t work directly with raw image data. It uses an autoencoder to compress images into a mathematical structure called a latent space. That compression process removes unnecessary visual data, which reduces the amount of information Flux.2 has to process and thereby speeds up inference.

Black Forest Labs has released an open-source version of Flux.2 on GitHub. It also provides two proprietary editions that offer additional features. FLUX.2 [pro] is designed to optimize the quality of outputted image, while FLUX.2 [flex] enables developers to trade off some output quality for faster processing times.

The company previously released two model families called FLUX.1 Kontext and Flux1.1 Pro. The former models are optimized to provide output customizability, while the latter algorithm series is better at generating realistic images. The company disclosed today that its image generators have been adopted by millions of users to date.

Black Forest Labs will use its new capital to finance engineering initiatives. The company plans to build “models that unify visual perception, generation, memory, and reasoning.” 

Its focus on equipping its algorithms with reasoning features suggest the company may be eyeing the robotics market. Robots require the ability to analyze visual data from sensors and make complex navigation decisions based on that data. If the company were to enter that market, it would face competition from fellow image generator developers such as Luma AI Inc., which closed a $900 million round last month.

Photo: Unsplash