FLUX 3 Image Klein Accelerates Image Generation on NVIDIA Blackwell GPUs

News

FLUX 3 Image Klein, optimized for NVIDIA Blackwell GPUs, delivers up to 4.45x faster image generation with TensorRT LLM VisualGen optimizations.

Black Forest Labs (BFL) has introduced FLUX 3 Image Klein, a cutting-edge image generation and editing model optimized for NVIDIA Blackwell GPUs. Leveraging TensorRT LLM VisualGen, the model achieves significant performance improvements, including up to 3.46x faster text-to-image (T2I) and 4.45x faster image-to-image (I2I) generation at 2048×2048 resolution compared to uncompiled BF16 inference on a single NVIDIA B200 GPU.

Optimizations for NVIDIA Blackwell GPUs

FLUX 3 Image Klein is designed to address the inference latency challenges inherent in diffusion models. The model incorporates Qwen3-VL-4B-Instruct for text conditioning, a diffusion transformer (DiT) for refining image latents, and an autoencoder for decoding and encoding image data. TensorRT LLM VisualGen introduces several optimizations, including fused kernels, multi-GPU parallelism, FP8/NVFP4 support, and caching, to enhance performance.

NVFP4, a quantized precision format supported by TensorRT LLM, plays a key role in accelerating inference. Tests on NVIDIA B200 GPUs showed that NVFP4 delivered up to 1.62x faster T2I and 1.58x faster I2I generation compared to BF16 precision. Additional improvements were achieved through optimized general matrix multiplication (GEMM) kernels and sparse attention mechanisms, such as Sol-Attention and VC-Attention FP8.

Performance Benchmarks

In benchmark tests, FLUX 3 Image Klein Base, which uses 50 denoising steps with a classifier-free guidance (CFG) scale of 3.5, demonstrated a total latency of 14.70 seconds for T2I tasks, with denoising accounting for 96.8% of the time. The distilled Klein variant, which uses 16 denoising steps without CFG, reduced total latency to 2.71 seconds, with denoising comprising 84.0% of the time. For I2I tasks, the Base model required 40.29 seconds, while the distilled variant completed the task in 7.48 seconds.

Quantized attention methods further reduced latency. VC-Attention FP8 achieved speedups of 1.18x for the Base model and 1.16x for the distilled variant, with minimal perceptual differences in output quality as measured by LPIPS scores.

Deployment and Use Cases

FLUX 3 Image Klein supports a wide range of applications, including text-to-image generation, image editing with up to 10 reference images, and layout-based instructions. The model is available under the FLUX Kommunity License, allowing free commercial use for individuals and businesses with annual revenue below $5 million. Larger organizations can access the model through commercial arrangements or an API priced at $0.009 per 1,000 images.

Developers can configure and deploy FLUX 3 Image Klein using YAML files provided in the TensorRT LLM VisualGen tutorial. These configurations enable users to take full advantage of the model’s optimizations, ensuring efficient and high-quality image generation for various workflows.