Black Forest Labs has released FLUX 3 Image Klein, a 5.6B parameter open weights model designed for versatile image generation tasks. Unlike many large-scale models, Klein is specifically optimized for consumer-grade GPUs, making it accessible for local deployment and private infrastructure. The model supports native 4K generation, pixel-perfect image editing, and multi-reference generation, where it can utilize up to 10 reference images to guide the output. It shares core capabilities with the full FLUX 3 Image model, including the use of bounding boxes to precisely control the placement of elements or to confine edits to specific areas of a frame.
Black Forest Labs has released FLUX 3 Image Klein, a 5.6B parameter open weights model designed for versatile image generation tasks.
The model's architecture leverages shared learnings across images, video, and audio to produce spatially grounded results. According to Black Forest Labs, these multi-modal signals help the model understand physical dynamics and interactions more effectively. In internal evaluations, FLUX 3 Image Klein outperformed larger open weights models and several closed models, including Ideogram V4.5 and FLUX.2 [pro], in categories such as text-to-image and multi-reference generation. Notably, it achieved a significant performance gain over its predecessor, FLUX.2 [klein], gaining 91 Elo points in text-to-image and 70 in image editing despite having 38% fewer parameters.
The model outperformed larger open weights models and several closed models, including Ideogram V4.5 and FLUX.2 [pro], in categories such as text-to-image and multi-reference generation.
Notably, it achieved a significant performance gain over its predecessor, FLUX.2 [klein], gaining 91 Elo points in text-to-image and 70 in image editing despite having 38% fewer parameters.
To balance speed and quality, the release includes both a base model for fine-tuning and a step-distilled model for low-latency production. The distilled version, which uses 16 steps, can generate a 1MP image in under 9 seconds on an NVIDIA GeForce RTX 4070. For high-performance environments, the model is optimized for NVIDIA's accelerated computing platform, capable of generating a 1MP image in under 0.6 seconds on an NVIDIA GB300 using NVFP4 quantization. While the weights are available under the FLUX Kommunity License—which allows free commercial use for individuals and companies with annual revenue under $5M—the model requires a specific prompt upsampler to function effectively, as short, typed prompts may result in generic outputs without the expanded scene captions provided by the upsampler.",
The distilled version, which uses 16 steps, can generate a 1MP image in under 9 seconds on an NVIDIA GeForce RTX 4070.