Reproduce “Adaptive Resource Management and Quality Control for Streaming Video Generation” 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
Yifei Xia, Hao Yuan, Suhan Ling, Haoran Sun, Hanke Zhang, Xupeng Miao, Fangcheng Fu, Bin Cui
SlackServe, a serving system for autoregressive diffusion transformers (AR-DiTs), enables real-time streaming video generation by adaptively managing computational resources and per-chunk fidelity based on a "playout slack" metric. The system achieves an average of 2.65x higher Continuous Play Ratio (QoE) and 3.39x lower Time to First Chunk (TTFC) compared to existing baselines, while maintaining high visual quality.
Key insights
- Playout slack serves as an effective, dynamic signal for guiding both resource management and quality control in real-time streaming video generation, enabling adaptive decisions to prevent user-perceptible stalls.
- Real-time streaming video generation allows for adjustable per-chunk fidelity based on current playout conditions, offering a deadline-sensitive quality-latency tradeoff distinct from offline generation.
- The impact of quality degradation from a low-precision historical Key-Value (KV) cache on the current chunk is limited, validating the feasibility of making per-chunk fidelity decisions with a degree of independence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.