Duke Universitys method is a little different. They developed pulse algorithm (photoupsampling via latent space exploration). Although it is SR super-resolution technology, it is not to fill pixels, but to create a large high-definition image, and then reduce the image resolution and compare with the original image to find the image with the highest matching degree.
Pulse algorithm can raise the low resolution image of 16x16 pixels to 1024x1024 level in a few seconds, and the accuracy has increased by 64 times, while the previous AI algorithm has improved by no more than 8 times.
With such technology, pulse algorithm can even restore a lot of details, including facial pores, hair, etc., to achieve realistic picture effect, real AI to mosaic.