- Published on
Speed Up Image Generation with FLUX 3 Image Klein on NVIDIA Blackwell GPUs
- Authors
- Name
- DeepHorus
Black Forest Labs’ FLUX 3 Image Klein can now run faster on NVIDIA Blackwell GPUs through TensorRT LLM VisualGen, with combined optimizations delivering up to 3.46x faster text-to-image and 4.45x faster image-to-image generation at 2048×2048 on a single NVIDIA B200. The gains apply to the Base checkpoint and compare against uncompiled BF16 inference. The model is an image editing and generation system trained on NVIDIA GB200 NVL72 and optimized for NVIDIA GPUs.
FLUX 3 Image Klein uses Qwen3-VL-4B-Instruct for text conditioning, a diffusion transformer to refine image latents, and an autoencoder to decode latents and encode image-to-image references. It processes text, noisy target-image, and optional reference-image streams independently before combining them through joint attention. BFL offers Base and distilled checkpoints for different latency and fidelity needs.
TensorRT LLM VisualGen, a dedicated runtime within TensorRT LLM, supported BF16, FP8, and NVFP4 variants at launch. On Blackwell, NVFP4 delivered up to 1.62x faster text-to-image and 1.58x faster image-to-image generation than BF16. VisualGen then targets denoising, which dominates overall generation time on B200, by optimizing the diffusion transformer.
The runtime uses optimized GEMM kernels and epilogue fusion, low-precision and sparse attention, caching, and multi-GPU execution. It also uses FlashInfer PrimsTS FP8/NVFP4 GEMM kernels built with CUTLASS Primitives and Task Scheduling. These kernels are tuned for FLUX 3 tensor shapes and enabled by default.
Benchmarks at 2048×2048 used NVFP4, VANILLA attention, and torch.compile, averaging 10 requests after two warmups. Base uses 50 denoising steps with a CFG scale of 3.5; distilled Klein uses 16 steps without CFG. The results highlight how deployment settings can materially change latency.
For design tools and visual production workflows, the optimizations reduce sequential passes and attention costs. Developers can apply the reported YAML configurations and deployment instructions to their workloads. Latency remains critical for interactive production workflows.
