At the end of last week, the startup Black Forest Labs announced the start of its activities in the development of generative neural networks. At the same time, the company, created by people from Stability AI, introduced a family of models for generating images based on text descriptions called FLUX.1, which claim to be the best in their class.
Image source: FLUX.1
The launch of FLUX.1 comes about seven weeks after Stability AI released AI generator Stable Diffusion 3 Medium in mid-June, which received a lot of criticism due to the poor quality of creating images that contain people. Users actively shared on social networks the results of generating an algorithm with distorted limbs and bodies of people.
The launch of Stable Diffusion 3 Medium follows the departure of three key employees from Stability AI – Robin Rombach, Andreas Blattmann and Dominik Lorenz. It was they, along with Patrick Esser, who participated in the development of the first version of Stable Diffusion and has since worked on various AI algorithms, as well as other engineers, who founded the Black Forest Lab company.
At the moment, the startup has presented three models for generating text descriptions FLUX.1. The FLUX.1 pro AI model is the most powerful image generator available for commercial use via its API. At the same time, FLUX.1 dev was released, available for non-commercial use, as well as the lighter and faster FLUX.1 schnell (translated from German as “fast” or “swift”).
The developers claim that their AI models are superior to existing analogues, such as Midjourney and DALL-E, in a number of indicators, including the quality of the generated images and the accuracy of following the original description. In general, the results of FLUX.1 generation are comparable to what can be created using DALL-E 3 from OpenAI in terms of accuracy of following the description, and are close in photorealism to Midjourney 6. Moreover, the algorithm clearly works better compared to Stable Diffusion XL, the latest a major release from the team of these developers when they were still part of Stability AI, not counting Stable Diffusion XL Turbo.
Black Forest Lab models are built on a hybrid architecture that combines transformation and diffusion methods, scaling up to 12 billion parameters. It seems that this approach makes the FLUX.1 neural network capable of generating high-quality human hands, which was the weak point of many analogues already released to the market. However, the developers did not specify on what data the FLUX.1 models were trained.
Note that Black Forest Lab has already attracted funding in the amount of $31 million. Those who want to try FLUX.1 in action can use the services Fal.ai or Replicate.com, where they will have to pay money to work with the algorithms.