Video compression stops chasing pixel perfection, uses AI to guess the details instead

Researchers built a compression system that throws away most of a video's pixels, then uses machine learning to reconstruct what it thinks should be there. Instead of storing pixel data, it stores a text description, a low-quality skeleton of the video, and optional motion cues — then a diffusion model fills in the missing details.

This is a fundamental shift in how compression works: stop trying to preserve the original data, start preserving what humans actually notice. Traditional codecs fail at ultra-low bitrates because they optimize for pixel accuracy, which is worthless if the result looks worse to your eye. This approach throws away the pixel problem entirely — it sends semantic information (what's in the scene) instead of pixel information (exact color values), and lets a generative model do the dirty work of synthesis. The practical effect: the same video takes 2 to 10 times less data to transmit, which matters for real-time streaming on poor connections, autonomous vehicle telemetry, and surveillance feeds where bandwidth is the bottleneck. The tradeoff is obvious — you're trusting an AI to hallucinate details instead of sending them — but the math says it works better than the alternative.