Currently developing a different take on this same concept, but using hand-modeled voxel shapes in place of sprites.

Haven't posted any significant updates in a while but progress is steady; just added cell-shading outlines to sprites that require it (can't translate black outlines into voxels otherwise) and improving the render pipeline to accommodate it.

http://n3s.io but more recent screenshots are at https://twitter.com/n3semu

This gave me a (stupid) idea. 2D to 3D seems possible to implement reasonably well with enough time and effort, given that it involves taking data and extrapolating it into another dimension -- for example, giving depth to sprites, etc. But what about the other way around? Taking 3D games (to keep things simple-ish, say from one particular platform) and automatically mapping them to 2D games. This sounds kind of absurd. Instead of instrumenting 2D game data with a little something extra, you'd have to compress/project the 3D data down into two dimensions -- in a way, stripping down the information contained in the original game. This is almost definitely impossible to pull off automatically for all games on a particular (3D) platform, at least in a way that isn't 100% awful to play. It's an interesting thought experiment to think of how it might be done, though.

seems similar to https://www.cs.cmu.edu/~tom7/zelda/zelda.pdf