Pulse: Self-Supervised Photo Upsampling via Latent Space Exploration (github.com)

This animation gives a good overview of what it is doing: https://raw.githubusercontent.com/adamian98/pulse/master/rea...

Basically it generates an upsampled image using a generative network, then downsamples it again to see how different it is from the original. It then tweaks the upsampled image (by moving around in the latent space of the generative model) until the downsampled version matches the original as closely as possible.

Seems like it works really well, but only for faces. I wonder if the same technique could be applied to a broader domain of images?

The actual paper can be found at https://arxiv.org/abs/2003.03808