FLUX.3 Image Klein lands in Diffusers: 5.6B open weights with bounding-box layout control
Serguey Asael Shinder2 min readHugging Face announced on 15 October 2026 that FLUX.3 Image Klein, the latest open image generation model from Black Forest Labs, runs in the Diffusers library from day one. The model card describes it as 5.6B open weights. One model handles text-to-image, image-to-image with several reference images, and something new for the family: layout control with bounding boxes written into the prompt.
Two checkpoints. Both are in the FLUX.3 Image Klein collection on the Hub:
black-forest-labs/flux-3-image-klein— timestep-distilled for fast inference, 16 steps, no classifier-free guidance;black-forest-labs/flux-3-image-klein-base— the base model, 50 steps with CFG, and the one intended for fine-tuning.
Layout control. After the caption, the prompt can carry a JSON list of regions, each with an id, a bbox and a desc, referred to in the caption as <id>. Boxes are written as [y_min, x_min, y_max, x_max] on a 0–1000 scale, so the same box works at any image size. For edits, each region gives a src_bbox in the reference and a tgt_bbox in the output; a region with identical boxes stays where it is. Several reference images are addressed as <ref_image_0>, <ref_image_1> and so on, in the order they are passed.
Speed and memory. End-to-end timings for the distilled checkpoint on a single H100 80GB, averaged over three runs: text-to-image at 1024 x 1024 in 1.93 s, image-to-image with one 1024 x 1024 reference in 3.87 s, and layout editing with a 1360 x 768 reference in 4.10 s. A reference roughly doubles the time because it adds 4,096 tokens to every transformer call. Peak memory after inference is 20.72 GiB in bf16, 16.27 GiB with the row-wise FP8 transformer on an H100 and 15.3 GiB with NVFP4 on an RTX 6000. The quantised files load with from_single_file; FP8 needs compute capability 8.9 or newer, NVFP4 needs Blackwell and FlashInfer, and the post notes that both lower memory without making generation faster.
Architecture. FLUX.3 is a latent flow model and reuses the FLUX.2 VAE: one token covers a 16 x 16 pixel patch, so a 1024 x 1024 image is 4,096 tokens. The text encoder is Qwen3-VL; the model reads hidden states from eight of its 36 layers and projects them to the transformer's width of 3,072. The transformer first runs 5 blocks per stream with separate weights for text, image and references, then 28 joint blocks with shared weights over the concatenated sequence. FLUX.2 let the streams attend to each other from the start; FLUX.3 keeps them apart first.
Training. Diffusers ships DreamBooth LoRA scripts for both text-to-image and instruction-based image editing, with an example that learns a tarot-card style. The post does not state the licence terms.
Source: Hugging Face Blog, "Welcome FLUX.3 Image Klein in Diffusers" — https://huggingface.co/blog/flux-3