SDXL Turbo Inference Speed Improved by Up to 229%

This introduces an optimization method leveraging the CUDA and TensorRT execution providers of ONNX Runtime to speed up inference for the SD Turbo and SDXL Turbo models.

In tests conducted on an NVIDIA GPU environment, the SDXL Turbo model recorded a throughput improvement of up to 229%, and the SD Turbo model up to 120%. ONNX Runtime outperformed PyTorch across all combinations of batch sizes and step counts.

The optimized models, generated through the model optimization tool Microsoft Olive, are readily available on Hugging Face. In addition, thanks to ONNX Runtime's cross-platform support, Stable Diffusion models can be run efficiently not only in Python but also in C# and Java environments.