HACKER Q&A
📣 graderjs

Deepfake Video Compression?


I was thinking that rather than have a video of Michael Jackson reading from Hamlet (assuming such thing existed anyway), couldn't we have a video that comprised of: AI likeness of MJ, text of Hamlet scene, other instructions (lighting, etc), and the Deepfake model can recreate it?

Is this a viable future video compression method? Also, doesn't it violate or potentially violate some fundamental Shannon limits of entropy / compression / information content regarding the content? If no, how does it get around them?

Bigger picture...what's the relationship between Kolmogorov Complexity / Minimum Description Length, and Shannon Entropy fundamental limits on compression? It would seem there is a paradox there on one level, how is that paradox usually resolved conceptually?

  👤 pizza Accepted Answer ✓
Nvidia already has a product that does this. It’s no more a violation of Shannon theory than neural audio codecs. All of these optimal compression models produce a string of length -sum(p log p). If you can come up with a better representation of likely events, that implies stronger assumptions on what the state space can actually be occupied by, and so you reduce the surprisal of individual observations, thereby reducing code length.