AMD GPU Guide

January 25, 2023 ยท View on GitHub

SD GUI comes with an implementation that enables hardware-accelerated AI image generation on all GPUs that support DirectML.

Note: The AMD implementation uses an entirely different codebase from the default Pytorch implementation and thus lacks certain features, such as tileable image generation or custom samplers.

Prerequisites

  • At least one Pytorch (ckpt) model file (included with the regular 3 GB SD GUI download)
  • A graphics card that supports DirectML (e.g. AMD 6XXX series)

Guide

  1. Open the Settings (F12) and set Image Generation Implementation to Stable Diffusion (ONNX - DirectML - For AMD GPUs).
  2. (Skip to #5 if you already have an ONNX model) Click the wrench button in the main window and click Convert Models.
  3. Select your model file to convert (e.g. sd-v1-5-fp16.ckpt).
  4. Set the Model Output Format to Diffusers ONNX (Folder), click Convert, and wait for it to finish.
  5. Select the ONNX model either in the settings or using the quick-switcher (Ctrl+M).
  6. Set your prompt and settings and click Generate! to generate images.

Alternatively, download and extract the already converted model from here.