I just can't understand how it manages to make this image http://www.pics.rs/i/2cuoM look like this image http://www.pics.rs/i/IEz0N, especially the last line of letters.

A lot of things happened since then. References up to spring 2014 are e.g. in our own work to super-resolve arbitrary sized images with convolutional en/decoders: Super-Resolution with Fast Approximate Convolutional Sparse Coding, http://brml.org/uploads/tx_sibibtex/281.pdf (which still has a lot of possibilities to be extended (e.g. color) and improved upon).

Interesting to see what some of them did since https://sites.google.com/site/dglasner/

Pretty original research.

So, if I understand from the abstract, for every pixel they look through the image for similar "pixel neighbourhoods" of the same or varying size, and collect them into a kind of database of examples to be used for scaling up that pixel. Pretty cool idea!

This is the algorithm they use in all those movies. "Enhance!"

I wonder if the same techniques could be used for compression as well.

Is there an implementation of this available?

To clarify my understanding of this post, would it be possible using this method to: 1. start with a high resolution image 2. create a low resolution version 3. using the low resolution version produce a high resolution version that looks good

It has a visual style that looks like a "photorealism" painting. Quite amazing. https://en.wikipedia.org/wiki/Photorealism

Original code @ https://github.com/stefanv/supreme (repost from downthread). Looks like it's BSD licensed and in Python.

Zoom and enhance

Discussed in 2012: https://news.ycombinator.com/item?id=4241266.