The encoding time is still a bit slow, but a new paper proves what many observers have been expecting from current AI algorithms for some time. In this experiment, a codec based on neural networks beats all previously known codecs in terms of image quality and efficiency. Simply explained, a sequence of frames is "understood" in its structure, which allows their reproduction without the usual artifacts. Instead, the finest details look more like painted with a watercolor brush:
If objective measurements such as those of structured similarity (MS-SSIM) are used to assess the quality, the new codec in the paper already wins hands down. However, it remains to be seen whether this would also be possible with other test videos or with special problems such as fast pans.
In any case, this seems to prove that algorithms with artificial intelligence will allow even lower compression rates for distribution.