Reproduce “Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation” 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
Hmrishav Bandyopadhyay, Xuanchi Ren, Zijian Huang, Jay Zhangjie Wu, Tianshi Cao, Ruilong Li, Bryan Chu, Sanja Fidler and 2 more
The authors introduce Causal Masked Distillation (CMD) for interactive autoregressive video generation, training a causal teacher to score targets using only past frames and available controls, with Prefix Scoring using the student’s generated history. On short-video VBench-I2V, Base CMD raises the benchmark’s overall aggregate Total from 82.65 to 88.33; long-video comparisons retain tradeoffs in temporal consistency and motion.
Key insights
- Directly aligning the teacher's information set with the causal constraints of the student during distillation is fundamental for robust autoregressive video generation.
- A unified causal framework, from teacher training and student initialization to inference, simplifies the distillation process and improves overall consistency.
- Regularizing the teacher's perception of the student's generated history via Prefix Corruption is crucial for stable learning, especially when the student's early outputs may contain artifacts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.