I've found the most useful aspect of upscaling is creating faces for people in scenes that are just a small bunch of messy pixels at 1024K. I wonder if this approach would work for that. Great videos by the way!
The Turbo scheduler from Stability seems to do a marginally better job than Karras on some generations. I am guessing due to its connection to the model. (The difference is very small so maybe my imagination!) It also shows how a sampler is assembled rather than having it all boxed up in one node.
