Generative video compression trades decoder compute for bandwidth savings
Information Capacity of Generative Video Compression: Quantifying the Rate-Compute Exchange at Identical Quality
Computer Vision and Pattern Recognition
Summary
Communicating videos usually means sending lots of data, but this paper studies a method that sends compact tokens and uses a powerful video decoder to create the video instead. The authors model how video quality depends on both the amount of data sent and the decoder’s computing power. They introduce a measure called information capacity to show how efficiently extra computing power saves bandwidth without losing quality. Their results show that bigger decoders can save more bandwidth for the same video quality, but this trade-off varies depending on the video content.
What this means in practice
- •For video streaming engineers: Choose decoder designs that optimally trade computation for bandwidth savings in video streaming systems to reduce network load without degrading quality.
- •For mobile device hardware teams: Decide when it is worthwhile to increase local decoding compute to save on mobile data usage for video playback.
Authors
Cheng Yuan, Jiawei Shao, Xuelong Li
Abstract
Under the AI Flow framework, communication networks distribute intelligence across devices, edge servers, and clouds, and computation at the receiver becomes a resource that can substitute for transmitted bits. Generative video compression (GVC) embodies this exchange by sending compact tokens with ultra-low bitrate and letting a generative decoder synthesize the video, yet how much bandwidth savings a unit of decoder compute actually achieves has never been quantified. To fill this vacancy, we model reconstruction quality as a two-factor power law in data rate and decoder compute, which fits measured DISTS of two GVC decoders with a mean error below 3%, and define the information capacity (IC) as the negative logarithmic slope along an iso-quality contour, namely the fraction of rate saved per fractional increase in compute at identical quality. IC is dimensionless and unit-invariant, thus enabling an architecture-agnostic comparison. It forms a field over the operating plane, locating where additional denoising steps are worth their cost. Across five datasets, the 14B decoder trades more compute for fewer rate about ten times more efficiently than the 1.3B decoder. IC also varies significantly across datasets, indicating imbalanced performance on the rate-compute trade-off in GVC methods.