Black Forest Labs released FLUX.3 Image Klein, and Hugging Face says it runs in the Diffusers library from day one.
What is new
One model covers text-to-image and multi-reference image-to-image editing. The headline addition is layout control: users write bounding boxes into the prompt to say where each part of an image should go, or which parts of a reference image to move.
Two open checkpoints are published. The time-step-distilled version defaults to 16 steps without classifier-free guidance, and the base model uses 50 steps with it. Per Hugging Face, the distilled checkpoint produces a 1024 × 1024 image in about two seconds on an H100. NVFP4 and FP8 weights are available for smaller GPUs, and LoRA training scripts cover both generation and editing. The text encoder is Qwen3-VL.
Why it matters
Layout control gives creators a more predictable way to compose scenes than prompt wording alone, and open weights let developers fine-tune the model themselves.