Researchers from ETH Zurich and Google have published a very interesting model for neural compression. Under the project title "High-Fidelity Generative Image Compression" or short HiFiC, a GAN (Generative Adversarial Network) with (pre)learned compression is hidden. The exciting thing is that on the project website you can try out how well the GAN compares to conventional compression methods in a "wipe comparison".
In fact, the demonstrated results are astonishing at particularly low data rates. While JPEG only shows blocks, HiFiC can still come up with amazing details.
But if you look closely compared to the original, you will see that some of the astonishing details just don't match the details in the original. Instead, you can see structures that typically fit well into the picture at this point.
We already mentioned it in the past and here you can try it out yourself. Many beautiful, highly compressed images that we will see in the future do not represent the reality we recorded, but only a typical approximation of it.
Welcome to the age of Computational Photography...
;)