The world is being quietly rearranged by people who write very long documents.


The title they went with Generation Is Compression: Zero-Shot Video Coding via Stochastic Rectified Flow Noisy translates that to

Video can be compressed by treating generation itself as the codec


Researchers have figured out how to use AI video generation models directly as compression tools instead of bolting them onto traditional codecs—the bitstream directly controls what the AI generates at each step. This means videos could be transmitted at much lower data rates while maintaining quality, since the AI model itself becomes the compression method rather than an add-on.
If this scales to real-world video systems, it could dramatically reduce bandwidth requirements for video streaming, which currently consumes about 80% of all internet traffic, by letting the receiving end reconstruct video using its own AI model rather than transmitting every pixel.

If you insist
Read the original →