LSVC: Enhancing Live Video Streaming with A Parallelized Learned Video Codec


Professor Klara Nahrstedt: Principle Investigator
Bo Chen: Ph.D. Student
Zhisheng Yan: Assistant Professor (George Mason University)
Yinjie Zhang: Ph.D. Student
Zhe Yang: Ph.D. Student


  • Fall 2021-Present

Project Description

Video codecs play an important role in live video streaming, accounting for a significant portion of today’s Internet traffic. Typically, live streaming systems are built with traditional codecs like AVC, HEVC, and MPEG, which have achieved great success during the past decades. In recent years, learned video codecs have been getting more attention by outperforming traditional video codecs in terms of coding efficiency. Nevertheless, the encoding speed of existing learned video codecs is far from supporting live streaming due to the slow forward propagation problem.
We propose LSVC, a parallelized learned video codec with an encoding speed that supports live streaming and state-of-the-art coding efficiency, to address this issue. LSVC accelerates encoding with a novel tree-based compression scheme, allowing parallelization in compression and trading GPU utilization for encoding speed. LSVC further improves its coding efficiency with a space-time-aware entropy module that fully exploits spatiotemporal features during video compression. LSVC outperforms H.265 by saving 8% bandwidth consumption in live streaming while improving the video quality by 0.4 dB in Peak-Signal-to-Noise-Ratio. LSVC achieves an encoding speed of 30 fps, outperforming DVC and RLVC, state-of-the-art learned video codecs, by 20% and 40%.

Funding Agencies

This project is supported by

  • The NSF Clowder Data Management funding (2022-2023).
  • The Army Research Lab funding (Until July 2022).


Bo Chen, Zhisheng Yan, Yinjie Zhang, Zhe Yang, and Klara Nahrstedt, “Lifter: Unleash learned codecs in video streaming with loose frame referencing” 21st USENIX Symposium on Networked Systems Design and Implementation (to appear).