I'm intrigued by the idea of deep learning systems 'hallucinating' high resolution images from low res. It makes sense: if it's a photo of known type (e.g. people, birds) you have a lot of additional data to interpolate intelligently.