Anyone who has used an image generator knows the frustration: the picture is good, but the dog is on the wrong side and the sign is in the middle when you wanted it in the corner. Black Forest Labs' FLUX.3 Image Klein, now supported in Hugging Face's Diffusers library from day one, adds a direct fix. You can write bounding boxes into the prompt to say where each part of the image should go, and move elements taken from a reference image.
The model handles both text-to-image and editing with several reference images in one checkpoint. Layout control works through a JSON list of regions in the prompt, each with an identifier, a box and a short description that the main caption refers to. For editing, the prompt describes the image you want with the change included, and reference images are cited in order. Black Forest Labs recommends full scene descriptions covering subject, setting, light, composition and style. The distilled checkpoint produces a 1024 x 1024 image in 16 steps by default, in about 2 seconds on an Nvidia H100, while a base checkpoint runs 50 steps with guidance and is meant as a starting point for fine-tuning.
Practical touches matter for smaller studios. Besides BF16 weights, Black Forest Labs ships rowwise FP8 and NVFP4 versions of the transformer for GPUs with less memory, and Diffusers includes DreamBooth LoRA training scripts for both generation and editing, so teams can teach the model a house style or a product. According to our review, the model has about 5.6 billion parameters and uses a Qwen3-VL-4B text encoder, and Black Forest Labs' own API charges from $0.008 per image at 768 pixels to $0.036 at 4K.
Placement control has been one of the weakest points of image models, which usually treat composition as a suggestion. Getting it in an open-weight model that runs on modest hardware is useful for designers, advertisers and anyone producing consistent layouts at volume, and Diffusers support means it slots into existing pipelines.
Check the licence before using it commercially. Our review notes that the weights are free only for organisations with annual revenue under $5 million; larger companies need a separate arrangement or the paid API. The speed figure is for an H100 at the default resolution, so smaller GPUs and larger images will be slower. And the Diffusers post shows what layout control can do in curated examples; how reliably it holds boxes in crowded scenes is something to test on your own prompts.