Black Forest Labs releases Flux 2 image models with open weights and high-res features

Black Forest Labs has launched FLUX.2, a new lineup of image generation models entering direct competition with offerings like the recently launched Google's Nano Banana Pro. The series enables image outputs up to four megapixels, supporting both generation and editing at high resolutions. While multi-image workflows are a focus, Flux 2 can process up to ten reference images simultaneously to ensure consistent characters and styles across outputs.

Text rendering has been reworked to improve tasks that depend on precise typography, UI mockups, and structured layouts. Prompt adherence has also improved, especially when handling complex instructions or dense visual compositions. Technically, Flux 2 uses a hybrid architecture combining the Mistral 3 24B vision language model with a Rectified Flow Transformer for accurate layout. A variational autoencoder encoder is included, allowing image data to be saved and restored efficiently with minimal quality loss.

The family includes the pro, flex, dev, and klein versions. The Pro model delivers top quality through the BFL Playground and BFL API, while the dev model offers open weights, a 32 billion parameter scale, and downloads hosted on Hugging Face with reference code on GitHub. Builds optimized with NVIDIA and ComfyUI run on consumer GPUs such as GeForce RTX cards, and klein will be released under Apache 2.0 with a beta already open.