Black Forest Labs has released FLUX.3 Image Klein, a new open-source image generation model available immediately in Hugging Face's Diffusers library. The model advances text-to-image and multi-reference image-to-image capabilities with a notable new feature: layout control, allowing users to specify exact positioning of image elements using bounding boxes written directly into prompts.
The release includes two model checkpoints—a distilled version optimized for speed and a base model for higher quality output. The distilled checkpoint generates 1024×1024 images in approximately 2 seconds on an H100 GPU without requiring classifier-free guidance, making it practical for real-time applications. The model supports quantized variants (NVFP4 and FP8) to run on smaller GPUs, alongside LoRA fine-tuning scripts for customization.
FLUX.3 Image Klein represents a significant step forward in open-source image generation, combining speed, quality, and unprecedented control over spatial composition. The modular Diffusers architecture allows developers to customize pipelines by adding or removing components, positioning the model as both a production tool and research platform for the AI community.
Key Points
Layout control feature lets users specify element positioning via bounding boxes in text prompts
Distilled checkpoint achieves 1024×1024 generation in ~2 seconds on H100 without classifier guidance
Both open checkpoints available in Diffusers immediately with LoRA training support
Quantized variants enable deployment on memory-constrained GPUs