NVIDIA says its TensorRT LLM VisualGen runtime makes FLUX 3 Image Klein, the Black Forest Labs model, up to 3.46 times faster at text-to-image generation on a single B200, cutting a 2048×2048 image from 22.36 to 6.46 seconds. The figure comes from a technical blog post published on November 6 and is measured against uncompiled BF16 inference, so it is a stack of optimizations rather than one switch.
Denoising takes about 97% of the Base model’s time, so nearly every gain targets that step
According to NVIDIA, its breakdown for the Base model at 2048×2048 on a B200 shows 14.24 of the 14.70 seconds of a text-to-image run spent in denoising, with NVFP4 weights and compilation already applied. The 16-step distilled model spends 84% of its 2.71 seconds there. Each optimization is aimed at that loop: NVFP4 weights alone are put at up to 1.62 times faster than BF16, the Cache-DiT step cache at 1.94 times for Base, and Sol-Attention sparse attention at 1.24 times, or 1.30 times with fewer dense steps.
On one B200, NVIDIA reports, compilation, NVFP4, Sol-Attention and Cache-DiT together give the 3.46 times for text-to-image and 4.45 times for image-to-image, from 54.64 to 12.28 seconds. The distilled model, which already runs 16 steps instead of 50, reaches up to 2.34 times.
Part of that speed comes from reusing and skipping work, and NVIDIA scores the cost as image drift
Cache-DiT reuses intermediate transformer outputs between denoising steps, and sparse attention skips part of the computation. NVIDIA measures what that does to the picture with LPIPS, a perceptual distance in which lower means closer to the baseline output. In its stacked text-to-image example the score rises from 0 for compiled BF16 to 0.044 with NVFP4, 0.084 with Sol-Attention and 0.095 with Cache-DiT. The post shows example images at each stage, and the only quality figures it reports are LPIPS distances to its own baseline output.
The results come from data center Blackwell parts, and the RTX PRO 6000 gain is smaller at the low end
The post says the optimizations and results focus on Blackwell data center GPUs and the RTX PRO 6000 Blackwell Server Edition. On that card a smaller stack of compilation, NVFP4 and Cache-DiT delivers 1.73 to 3.25 times, against 3.46 times on the B200. In a separate test, spreading the work over eight B200s in a DGX system cuts Base text-to-image latency from 14.90 seconds on one GPU to 3.00, and the distilled model from 2.73 to 1.08 seconds, where roughly half the time falls outside denoising and limits the gain.
There are no GeForce numbers in the post, so the ratios may not carry over to a consumer RTX card. The route in is NVIDIA’s prebuilt TensorRT LLM container, which can run FLUX 3 offline or serve it through an OpenAI-compatible image generation endpoint.