StreamDiffusion achieves 106 fps txt2img and 93 fps img2img on RTX 4090 with SD-Turbo plus TensorRT

On a reference system (RTX 4090 GPU, Core i9-13900K CPU, Ubuntu 22.04) with SD-Turbo at 1 denoising step and TensorRT acceleration, StreamDiffusion benchmarks at 106.16 fps text-to-image and 93.897 fps image-to-image. With LCM-LoRA + KohakuV2 at 4 denoising steps the rates are 38.023 fps and 37.133 fps respectively. The paper reports a 59.56× throughput improvement over Diffusers’ AutoPipeline baseline. These numbers set expectations for live performance use: 90+ fps provides smooth real-time video; 38 fps is still sufficient for video-rate output. Exact numbers depend on resolution (default 512×512), batch size, and acceleration backend.

Examples

Minimum viable setup for 60fps live performance: SD-Turbo + TensorRT + t_index_list=[0] + 512×512. LCM-LoRA models at 4 steps hover around 37–38 fps—suitable for 30fps output with headroom.

Assessment

Which two variables (model and acceleration mode) most strongly determine whether StreamDiffusion can maintain >60 fps? Predict fps order for: (a) SD-Turbo + TensorRT, (b) KohakuV2 + LCM-LoRA + xformers, (c) KohakuV2 + LCM-LoRA + TensorRT.