Nvidia reports faster FLUX 3 image generation with VisualGen optimizations

Nvidia reported that combined runtime optimizations accelerated FLUX 3 Image Klein Base image generation on a single B200 GPU. At 2048×2048 resolution, text-to-image latency fell from 22.36 to 6.46 seconds, while image-to-image latency fell from 54.64 to 12.28 seconds compared with uncompiled BF16 inference. Its TensorRT LLM VisualGen tutorial provides deployment instructions combining compilation, lower-precision computation, sparse attention and reuse of intermediate model outputs. These methods reduce computation during the repeated steps that generate an image. Sparse attention and caching affect output quality, and the reported latencies exclude the separate prompt upsampling process.