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
- Open the Settings (F12) and set
Image Generation ImplementationtoStable Diffusion (ONNX - DirectML - For AMD GPUs). - (Skip to #5 if you already have an ONNX model) Click the wrench button in the main window and click
Convert Models. - Select your model file to convert (e.g.
sd-v1-5-fp16.ckpt). - Set the
Model Output FormattoDiffusers ONNX (Folder), clickConvert, and wait for it to finish. - Select the ONNX model either in the settings or using the quick-switcher (Ctrl+M).
- Set your prompt and settings and click
Generate!to generate images.
Alternatively, download and extract the already converted model from here.