Black Forest Labs releases open FLUX.3 Image Klein models with day-one Diffusers support and bounding-box layout control

Black Forest Labs released two open FLUX.3 Image Klein checkpoints on Hugging Face: a distilled one (16 steps, no CFG) and a base one (50 steps, with CFG). Both run in Diffusers from day one as a Modular Diffusers pipeline. A single model handles text-to-image, multi-reference image-to-image and bounding-box layout control, using a Qwen3-VL text encoder. BFL also ships single-file bf16, FP8 and NVFP4 transformer weights, plus LoRA training scripts for text-to-image and editing.

Developers and researchers working with open image models get a downloadable model with a practical new control (placing objects by bounding box) and smaller quantised variants that may fit more modest GPUs. The LoRA scripts lower the cost of fine-tuning for custom styles or products. The impact is mainly incremental for teams already using Diffusers. Whether it is worth switching depends on licence terms, which the coverage does not state, and on quality compared with current open alternatives.

The "about 2 seconds for a 1024x1024 image on an H100" figure is self-reported, with no independent benchmark and no detail on settings. The post was co-written with Hugging Face staff and reads partly as promotion, and claims such as "easy to fine-tune" are untested. Missing context includes the licence, output quality against rival open models, and real memory needs for the quantised variants.

1. 01The licence on the model cards: whether commercial use is permitted, and whether the distilled and base checkpoints differ.
2. 02Independent benchmarks and community tests of speed and image quality, including reproduction of the 2-second H100 figure and results for FP8 and NVFP4 on consumer GPUs.
3. 03Adoption within weeks: Hugging Face download counts, community LoRA releases, and integrations in tools such as ComfyUI, plus how well the bounding-box control holds up in real use.