Black Forest Labs has introduced FLUX.3 Image Klein, a latent-space flow model for image generation that is available in Hugging Face Diffusers from day one. According to Hugging Face, the model supports text-to-image generation as well as multi-reference image-to-image workflows, introducing a new layout control feature. Users can write bounding boxes into prompts to dictate the positioning of image parts or to move sections of a reference image. Both open checkpoints, including the distilled version that generates a 1024 by 1024 image in 16 steps, are available in the FLUX.3 Image Klein collection. The architecture employs the same VAE architecture as Flux.2, compressing images into tokens, while utilizing Qwen3-VL as its text encoder. FLUX.3 reads hidden states from eight layers of the 36-layer text encoder and stacks them before projection. The model separates text, image, and reference tokens into distinct streams initially before merging them. Additionally, Black Forest Labs provides the transformer as a single file in bf16 and quantized formats, alongside DreamBooth LoRA training scripts for custom model adaptations.
Welcome FLUX.3 Image Klein in Diffusers
Black Forest Labs has released FLUX.3 Image Klein, supporting text-to-image and multi-reference image-to-image tasks with new layout control capabilities. The model is integrated into Hugging Face Diffusers from day one. It utilizes a latent-space flow architecture, a Qwen3-VL text encoder, and the Flux.2 VAE.
Why it matters
The release of FLUX.3 Image Klein provides the AI community with advanced open checkpoints supporting precise layout control and multi-reference editing directly within Diffusers.