This method proposed by AMD is great, but only certain models are supported.
A better method that supports all models, and increases compute speed is this. When running webui.bat, you use arguments
- "webui.bat --onnx --backend directml" for ONNYX,but include this rather:
- "webui.bat --backend directml --opt-sub-quad-attention"
This will increase compute dramatically for any traditional checkpoints you use, such as ReV_Animated
At Least, this has been my experience with a Radeon RX 6800