Reproduce “Generative Transmission: Rethinking Computation, Bandwidth, and Memory in Communication” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiangyu Chen, Jixiang Luo, Yuankai Fan, Haibin Huang, Chi Zhang, Xuelong Li
A new framework, Generative Transmission (GenTrans), rethinks video communication for ultra-low-bandwidth and weak-network conditions by jointly optimizing bandwidth, computation, and memory. It demonstrates superior perceptual quality at extreme bitrates, achieves a 1.5x decoding speedup, and maintains robust visual content even with up to 50% packet loss.
Key insights
- Reframing video communication as a joint optimization of bandwidth, computation, and memory, rather than just rate-distortion, enables more efficient and robust systems for weak networks.
- Exploiting cross-clip temporal redundancy through generative memory offers substantial bandwidth savings by re-using stable scene content, especially for low-motion scenarios.
- Receiver-side computational overhead in generative video can be effectively mitigated by reusing intermediate runtime states, significantly improving decoding efficiency and reducing latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.