Flux 3 image editing up to 4.45x faster… measured on a single B200

NVIDIA reported that optimizing FLUX 3 Image Klein Base cut 2048×2048 image editing time on a single B200 GPU from 54.64 seconds to 12.28 seconds, a 4.45x speedup versus unoptimized BF16 execution under the same conditions. The company said in a technical blog post that text-to-image generation on the same setup fell from 22.36 seconds to 6.46 seconds, a 3.46x improvement. The gains came from combining compilation, NVFP4 low-precision arithmetic, a PrimsTS GEMM matrix kernel, Sol-Attention and Cache-DiT rather than from NVFP4 alone. The results matter because they show how much inference cost and latency can be reduced on existing Blackwell hardware through software optimization, a key lever as image and video generation workloads scale. What to watch next is whether these speedups hold across other batch sizes and GPU configurations, and how quickly they are adopted in production pipelines.