Reproduce “MotionPhys: Detecting AI-Generated Videos via Physical Consistency of Optical-Flow Trajectories” 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
Haojin He, Hao Tan, Zichang Tan, Ajian Liu, Jun Wan
MotionPhys detects AI-generated videos by analyzing systematic inconsistencies in the physical properties of optical-flow trajectories, rather than relying on visual artifacts. The framework achieves 78.5% accuracy on GenBuster++ and an average AUC of 84.8% in cross-generator tests on AIGVDBench, surpassing several existing detection methods.
Key insights
- AI-generated videos, despite their visual plausibility, exhibit systematic statistical discrepancies and implicit deviations from real-world physical motion consistency in object trajectories.
- Fundamental physical properties of motion, such as geometric bending (curvature), directional evolution (angular velocity), and high-order smoothness (jerk), provide robust and generalizable cues for detecting synthetic videos.
- Combining multi-scale trajectory analysis with statistical aggregation effectively captures both short-lived local irregularities and accumulated inconsistencies in generated motion dynamics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.