NVIDIA reported a 7.48-second image edit at 2,048-by-2,048 resolution using FLUX 3 Image Klein's distilled model and TensorRT LLM VisualGen on one B200 GPU.
In a separate comparison, NVIDIA's combined optimizations made image-to-image generation on the Base model 4.45 times faster and text-to-image generation 3.46 times faster than uncompiled BF16 inference. The figures describe the reported setup and do not establish performance on other hardware or configurations.
For Base, NVIDIA used 50 denoising rounds and set classifier-free guidance to 3.5; text-to-image took 14.70 seconds, while editing took 40.29 seconds. Distilled ran 16 rounds without that guidance, completing the two tasks in 2.71 and 7.48 seconds.
Compared with BF16, NVFP4 reached speed gains of as much as 1.62 times for text-to-image tasks and 1.58 times for image edits. VC-Attention FP8 raised speed by 1.18 times on Base and 1.16 times on distilled, and NVIDIA used LPIPS measurements to assess visual differences.
The model's pipeline routes text prompts through Qwen3-VL-4B-Instruct, processes image representations with a diffusion transformer and uses an autoencoder to encode and decode images. NVIDIA lists fused kernels, multi-GPU processing, FP8 and NVFP4 support, and caching among VisualGen's optimization options.
Black Forest Labs announced the model's open-weight release on Oct. 15. NVIDIA's TensorRT LLM VisualGen tutorial gives developers YAML examples for deployment setup.
