I finally got around to creating a new image workflow to share. I've tried to hide a lot of the complexity in a subgraph and only expose the necessary parameters.
The workflow creates three images: the base generation, the upscaled and refined generation and an optional face enhancer using my GonzaLomo ZPop v4.0 model. I think it does result in a better image because SDXL faces are not perfect. The face detailer requires some special detection models which you can get here.
The workflow also has a simple IPAdapter integration. So it can take a reference image and incorporate that into the image generation.