Black Forest Labs Releases FLUX.3 Image Klein with Day-One Diffusers Support

Black Forest Labs released FLUX.3 Image Klein on October 15, 2026, with day-one support in Diffusers. The model combines text-to-image and multi-reference image-to-image in one architecture and adds layout control: users can include JSON bounding boxes in the prompt to specify where image regions should appear, or use src_bbox/tgt_bbox pairs to move content within a reference image.

Two open checkpoints are available. The distilled checkpoint (flux-3-image-klein) uses 16 steps without classifier-free guidance for faster inference, while the base checkpoint runs 50 steps with CFG and is intended for fine-tuning. According to the release, the distilled model generates a 1024×1024 image in about two seconds on a single H100. NVFP4 and FP8 quantized transformer weights can be loaded with from_single_file to reduce memory usage, and DreamBooth LoRA training scripts are provided for both text-to-image and image editing.

Under the hood, FLUX.3 is a latent-space flow-matching model. It uses the Qwen3-VL text encoder, reading hidden states from eight of its 36 layers, and the same VAE as Flux.2. The model ships as a Modular Diffusers pipeline, enabling users to inspect workflows, remove components such as the text encoder, or insert custom blocks into the denoising loop. A Hugging Face Space is available for online testing.