This isn't like most super-resolution techniques, that take multiple images and align them to a higher-resolution grid. https://en.wikipedia.org/wiki/Super-resolution#Geometrical_o... This is about having a model of what a face should look like, and fitting the data to it. It's like taking a handful of data points and plotting a curve, then interpolating/extrapolating other data points. Except instead of something simple like a bell curve, the model is of a human face! (That means you could feed it random data and it would still come up with a face - grotesquely distorted perhaps, but the model can only make faces.)