Black Forest Labs open-sources 6B FLUX.3 Image Klein on October 15, unifying text-to-image, multi-reference and layout control

Black Forest Labs on October 15 open-sourced the 6B image model FLUX.3 Image Klein via Hugging Face Diffusers, unifying text-to-image, multi-reference image-to-image and bounding-box layout control in one model; the layout control allows users to specify positions of image parts via bounding boxes or move elements from reference images to target positions. The model uses Flux.2's 8x compression VAE so each token covers 16x16 pixels; the text encoder Qwen3-VL takes hidden states from layers 4 to 32 in steps of 4 (8 layers), concatenated and projected to dimension 3072, giving each token a 20,480-dimensional representation. The pipeline runs in two stages: first each token type is refined independently through 5 blocks, then all tokens are concatenated as a text+image+reference sequence and processed jointly by 28 shared-weight blocks, with a 4D RoPE position encoding in (t, h, w, l). Checkpoints include black-forest-labs/flux-3-image-klein (distilled, 16 steps, no CFG) and black-forest-labs/flux-3-image-klein-base (base, 50 steps, CFG); NVFP4 and FP8 quantised weights are also available. Benchmarks show the distilled version running at around 1.93 s for 1024x1024 text-to-image, 3.87 s for 1024x1024 single-reference image-to-image and 4.10 s for 1360x768 layout editing on a single H100 80GB, peaking at about 23.8 GiB of VRAM. The model ships as a Modular Diffusers pipeline and includes DreamBooth LoRA fine-tuning scripts and an official Space demo.