Black Forest Labs announced the release of a new family of image generation models FLUX.2. The new model is focused on real creative tasks and allows for the creation of high-quality images with adherence to style, characters, and complex composition requirements. FLUX.2 supports working with up to 10 reference images simultaneously, ensuring consistency in style and detail of products or characters across different frames.
The model generates photorealistic images, infographics with complex typography, and interface layouts with readable text. Users can edit images with a resolution of up to 4 megapixels without losing quality and detail integrity. FLUX.2 accurately executes complex text instructions, adheres to brand guidelines, and correctly handles lighting, element layout, and logos.
Various options are available for developers: FLUX.2 [pro] for maximum quality and speed, FLUX.2 [flex] with the ability to control generation parameters, FLUX.2 [dev] with open weights for local deployment, and FLUX.2 [klein], open under the Apache 2.0 license, which will soon be available in open access. FLUX.2 [dev] can be downloaded from Hugging Face and run on consumer graphics cards, including GeForce RTX, thanks to collaboration with NVIDIA and ComfyUI.
All FLUX.2 models combine image generation and editing in a single architecture, using “latent flow matching” and the visual-language model Mistral-3 24B. This allows for stability in complex scenes, accurate text reproduction, and high realism. The new FLUX.2 autoencoder – VAE ensures a balance between quality, training speed, and compression degree, and is already available under an open license.
FLUX.2 is available through BFL Playground, the company’s API, and in partner services. The model opens up new possibilities for designers, developers, and teams who need control over image quality, generation speed, and cost.
