FLUX.3 Image Klein lands in Diffusers with layout control

FLUX.3 Image Klein arrives with an unusual combination: a single checkpoint that switches between generating an image from text, editing one or more reference images, and placing elements at specific coordinates. That last capability, called layout control, works by writing a JSON list of regions into the prompt, each with an id, a bbox, and a desc, on a normalized 0–1000 scale so the same box works at any image size. For edits, each region says where its content comes from in the reference and where it goes in the output, which is what lets a user move one object, such as a knitted knight, while pinning the rest of the scene.

The release also leans heavily on the tooling around it. The model ships as a Modular Diffusers pipeline, so developers can pop the text encoder out, encode a prompt once, free it, and generate many images from the cached embeddings, or insert their own step inside the denoising loop. Black Forest Labs provides the transformer as a single file in bf16 plus two quantized formats, fp8r and nvfp4, loadable with from_single_file for smaller GPUs, and Hugging Face supplies DreamBooth LoRA training scripts for both text-to-image and image-to-image tasks. Together these choices lower the barrier for teams that want to fine-tune on their own data or run on hardware that is not top of the line.