Black Forest Labs: AI startup for image and video from the Black Forest

They have already worked on the well-known image generator Stable Diffusion at Stability.ai, and now some developers have set up their own business and founded the start-up Black Forest Labs. The company is based in Wilmington, North Carolina, according to its legal notice. The start-up has already raised 31 million US dollars in a recent financing round, including from well-known tech investor Andreessen Horowitz.

Black Forest Labs announces on its website that it wants to develop state-of-the-art generative deep-learning models for images and video. "We believe generative AI will become a fundamental part of all future technologies." The wide availability of the models, which will also be published as open source, should allow everyone to benefit from the technology - and create more trust in safety.

They should be the best image and video generators

The start-up has already announced the Flux.1 image generator with the promise that the model family will push the boundaries of text-to-image synthesis. The "new state of the art" refers to "image details, prompt compliance, style diversity and scene complexity". There are three versions of the image generator. Flux.1 [pro] is the most powerful model and is accessible via API - as well as via the Replicate and fal.ai platforms. There are also customized company versions. Flux.1 [dev] is, as the name suggests, freely available for non-commercial purposes - for example via HuggingFace, Replicate and fal.ai. Flux.1 [fast] is a particularly fast model made for personal and local use. It is licensed under the Apache 2.0 license.

The models are based on a hybrid architecture of multimodal and parallel diffusion transformer blocks. A detailed tech report will be published soon. According to the company, Flux.1 [pro] and [dev] already outperform current models such as Midjourney v6.0, Dalle-E 3 from OpenAI and SD3-Ultra from Stable Diffusion in some benchmarks. These include visual quality, prompt compliance, typography, output diversity and size and aspect variability. The fast version of Flux.1 also outperforms models of the same class.

Black Forest Labs has already announced a video generator, which is also expected to be the best of its kind. The model will apparently be called SOTA - as stated on the homepage under "What's next" - and will be available to everyone. OpenAI's video generator is called Sora and, despite being announced six months ago, can only be used by selected testers.

The team behind Black Forest Labs left Stability.ai at the beginning of the year. In a seed financing round, i.e. a very early round, they have now already raised 31 million US dollars. Andreessen Horowitz is significantly involved, as are other smaller angel investors. In the announcement, the founders state that they want to bring state-of-the-art AI from Europe to the world.

(emw)