Reproduce “Chain of Event-Centric Causal Thought for Physically Plausible 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
Zixuan Wang, Yixin Hu, Haolan Wang, Feng Chen, Yan Liu, Wen Li, Yinjie Lei
A framework called Chain of Event-Centric Causal Thought (CoECT) is introduced for physically plausible video generation, which explicitly models complex physical phenomena as a sequence of causally ordered and dynamically evolving events. This approach demonstrated an 8.19% improvement in Physical Commonsense Alignment on PhyGenBench and a 3.4% gain in combined Semantic Adherence and Physical Commonsense on VideoPhy compared to prior methods.
Key insights
- Explicitly modeling complex physical phenomena as a sequence of causally ordered event units, grounded in physical formulas, is crucial for generating videos that accurately reflect real-world physics.
- Decomposing physical processes into fine-grained events with dynamically evolving physical conditions and scene graphs allows for the generation of realistic and causally coherent temporal progressions.
- Leveraging a dual-condition strategy through progressively revised semantic narratives and interactively synthesized, physically constrained keyframes effectively guides video diffusion models, ensuring both temporal consistency and visual plausibility across event transitions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.