FLUX.2: Image AI Model with Improved Text Rendering and World Knowledge
FLUX.2, the successor to the popular image AI model FLUX.1, offers improved text rendering, expanded world knowledge, and higher resolution.
(Image: Black Forest Labs)
With FLUX.2, the globally popular image AI model from Black Forest Labs, based in Freiburg, has now received a successor. Like its predecessor, FLUX.2 follows the "Open Core" approach. This means the model can be operated locally, and users can adapt it to their needs. FLUX.1 is considered one of the most popular open image models. The AI startup was founded just a year ago.
The new version adds capabilities that FLUX.1 was still missing. This includes, for example, better text rendering. The model is intended to output typography, infographics, memes, and user interface mockups with easily readable text. Those who want to generate something based on previous images can now incorporate up to ten reference images simultaneously. Previously, only a single one was possible. This is intended to ensure that characters, products, and styles remain consistent in subsequent image generations.
Expanded World Knowledge and Higher Resolution
The developers also promise sharper textures, more precise details, and stable lighting. This should make FLUX.2 more suitable for product images. In this context, the improved adherence to prompts should also be helpful. Complex, multi-part instructions are intended to be implemented more reliably.
Videos by heise
The new model also features expanded world knowledge. A better understanding of physics, lighting, and spatial relationships is intended to make the outputs look more realistic. Additionally, image editing is possible up to a resolution of 4 megapixels, and aspect ratios can be set more flexibly.
Various Model Variants
FLUX.2 is available in different model variants. FLUX.2 pro, with the highest quality, is available via API and partners. FLUX.2 flex allows developers full control over parameters but is also only usable via API and partners. FLUX.2 dev, with 32 billion parameters, can be downloaded as an open-weight model, among other places, at Hugging Face. Black Forest Labs has also announced FLUX.2 klein, an open-source model currently available in a closed beta. And there is also FLUX.2 VAE as a component, a new autoencoder serving as the basis for all FLUX.2 models.
(mki)
