This extension enables optimized execution of base Stable Diffusion models on Windows. Because DirectML runs across hardware, this means users can expect performance speed-ups on a broad range of accelerator hardware.

As a pre-requisite, the base models need to be optimized through Olive and added to the WebUI's model inventory, as described in the Setup section. The extension uses ONNX Runtime and DirectML to run inference against these models.

This preview extension offers DirectML support for compute-heavy uNet models in Stable Diffusion, similar to Automatic1111's sample TensorRT extension and NVIDIA's TensorRT extension.

Stable Diffusion versions 1.5, 2.0 and 2.1 are supported.