ML-based super resolution / upscaling

This has previously been proposed in one form for comments and voting here: