FLUX.3 Image Klein arrives as an open image model with bounding-box layout control
3-line summary
- Black Forest Labs' FLUX.3 Image Klein image model is out and works in Hugging Face Diffusers from day one.
- One model handles text-to-image, multi-reference editing and new layout control that places elements with bounding boxes.
- The fast version makes a 1024×1024 image in about 2 seconds on an H100, and compressed weights are available for smaller GPUs.
Hugging Face announced on October 15 that FLUX.3 Image Klein, Black Forest Labs' new image model, works in the Diffusers library from day one. Two checkpoints are out: a step-distilled version for fast generation (16 steps) and a base version for fine-tuning (50 steps).
The new feature is layout control. Add bounding boxes and descriptions for each region after the prompt, and each element is drawn in its place; you can also move just one part of a reference image while leaving the rest untouched. Several reference images can be combined too, as in "put this fox on that bench."
On an H100 GPU, a 1024×1024 image takes about 1.93 seconds, and a reference image roughly doubles that. FP8 and NVFP4 compressed weights help smaller GPUs, and LoRA training scripts let you teach the model a particular style.
You can try it without installing anything in a Hugging Face Space. Those speeds are for high-end server GPUs; consumer-PC speeds were not published. Check the license on the model page before using it commercially.
Why it matters
Getting the right thing in the right place has been one of the most frustrating parts of image generation. With coordinate-based layout control in an open model, producing repeated fixed-layout work such as banners and product pages should get easier.
Who it's for
Sources
- Welcome FLUX.3 Image Klein in Diffusers — Hugging Face
