You can't easily auto-tween a 3D rotation of a 2D character, for example.

I think the author's approach of "use upsizing algorithms as a basis for humans to remaster" is probably the best of both worlds.

Actually, you can. See the previous HN discussion of "Framefree", which decomposes an image into separately moving layers and uses a full mesh morph layer by layer to interpolate between frames. Some ultra-slow motion for sports is done that way.

That's pretty awesome! I stand by my point that it's not easy, but I'm amazed by how effective that is.