# FLUX 3 Image Klein Runs Image-to-Image Tasks Up to 4.45 Times Faster By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/51835 Published: 2026-11-08T21:54:10.000Z Updated: 2026-11-08T21:54:10.000Z Section: Technology > The 2,048-by-2,048 benchmark compared TensorRT LLM VisualGen with uncompiled BF16 inference on one NVIDIA B200 GPU. NVFP4 also improved generation speeds against BF16. FLUX 3 Image Klein produced 2,048-by-2,048 images up to 4.45 times faster for image-to-image tasks and 3.46 times faster for text-to-image on one NVIDIA B200 with TensorRT LLM VisualGen, versus uncompiled BF16 inference. The Base configuration used 50 denoising steps and a classifier-free guidance scale of 3.5. It completed text-to-image generation in 14.70 seconds and image-to-image generation in 40.29 seconds. The distilled version used 16 steps without classifier-free guidance, taking 2.71 seconds and 7.48 seconds, respectively. The inference system uses Qwen3-VL-4B-Instruct to condition prompts, a diffusion transformer to refine image data and an autoencoder to encode and decode it. TensorRT LLM VisualGen’s optimization set includes fused kernels, multi-GPU parallelism, FP8 and NVFP4 support, and caching. NVFP4, a lower-precision number format, generated text-to-image results up to 1.62 times faster and image-to-image results up to 1.58 times faster than BF16. VC-Attention FP8 delivered additional speedups of 1.18 times for the Base version and 1.16 times for the distilled version, with minimal perceptual differences in outputs measured by LPIPS scores. Developers can set up deployments with YAML files in the TensorRT LLM VisualGen tutorial. Black Forest Labs released FLUX 3 Image Klein on Oct. 15 with [commercial-use terms tied to revenue across affiliated entities](). ## Links in this article - [commercial-use terms tied to revenue across affiliated entities](https://www.tokenpost.com/news/technology/32877)