I cannot compare the performance with other cards, but it takes a few seconds for SDXL images (e.g. 1024x512) as long as it doesn’t run OOM.

I use a fork of the stable diffusion webui [0] which, for me, handled memory better. Setup was relatively easy: install the pytorch packages from the ROCm repo and it worked.

[0]: https://github.com/lllyasviel/stable-diffusion-webui-forge