Speed Up Image Generation with FLUX 3 Image Klein on NVIDIA Blackwell GPUs | NVIDIA Technical Blog - NVIDIA Developer

Deploying diffusion models in design tools and visual workflows requires addressing inference latency, especially when generating high-resolution images through models with billions of parameters. FLUX 3 Image Klein, an image editing and generation model from Black Forest Labs, has been trained on NVIDIA GB200 NVL72 and optimized specifically for NVIDIA GPUs.

Performance Improvements

In tests conducted on a single NVIDIA B200 GPU at 2048×2048 resolution, combined optimizations delivered up to 3.46x faster text-to-image (T2I) and 4.45x faster image-to-image (I2I) generation for FLUX 3 Image Klein Base compared with uncompiled BF16 inference. The tutorial includes deployment instructions and YAML configurations for applying these optimizations to various workloads.

Technical Optimizations

TensorRT LLM VisualGen, a dedicated runtime within TensorRT LLM, supports BF16, FP8, and NVFP4 variants. On Blackwell GPUs, NVFP4 delivered up to 1.62x faster T2I and 1.58x faster I2I generation than BF16. The optimizations focus on accelerating FLUX 3's DiT (diffusion transformer) component, which dominates overall generation time.

VisualGen accelerates NVFP4 DiT inference with optimized GEMM kernels and epilogue fusion, low-precision and sparse attention, caching, and multi-GPU execution. These improvements make the model more practical for real-time applications in design and visual production.