A new research pipeline named PRESLEY shows you can shrink video streaming bitrate by having a generative AI model guess at background details instead of transmitting them outright.
PRESLEY builds on an earlier system called ELVIS. Instead of deleting background blocks outright, it degrades them in place, guided by a mask that decides how much each block can afford to lose. A compact side channel tells the decoder how hard each block was hit. On the receiving end, a generative model reconstructs those blocks using visual cues sent alongside the video, rather than guessing blind. Tested across 13 rate ladders spanning multiple codecs and datasets, PRESLEY cut background bitrate by an average of 56.4% compared to its predecessor, and by up to 29.4% against standard, unmodified encoding in bandwidth-starved conditions, all while leaving the parts of the frame viewers actually look at completely untouched.
The bigger story is what the researchers found about the limits of this approach. Their own math is the interesting part: existing heuristics for picking which blocks to degrade already capture 83.3% of the savings a theoretically optimal selector could achieve, leaving only about 5% of total bitrate as remaining headroom on that front. The real unsolved problem is different: how badly restoration damages a block varies wildly, by 4.9 to 8.4 dB, though the researchers show that damage is predictable before the video is even sent.
This isn't a new codec so much as a bet that spare compute on your device is cheaper than spare bandwidth on the network, a trade only useful if viewers never notice the AI's background is a guess.