Black Forest Labs’ FLUX.3 Image Klein is available in Hugging Face Diffusers with text-to-image generation, multi-reference editing and explicit layout control in one pipeline. Developers can describe regions as bounding boxes inside a prompt to position new objects or move selected parts of a reference image.
The layout format uses normalized coordinates from 0 to 1,000, so the same box can be reused at different image sizes. For an edit, a prompt can identify both the source region and its destination while pinning the rest of the composition in place. Multiple source images can also be named separately and combined in one output.
Two open checkpoints are available. The distilled version generates at 1,024 by 1,024 pixels in 16 steps and took about two seconds on an H100 in Hugging Face’s example. The base model uses 50 steps with classifier-free guidance and is intended for fine-tuning. LoRA scripts cover both generation and image editing.
Hugging Face also provides BF16, row-wise FP8 and NVFP4 transformer weights. The quantized files reduce memory requirements, but actual speed and output quality will depend on the GPU and precision selected. The release uses a modular pipeline, allowing developers to replace or add individual components.