Reproduce “Grab-3D: Detecting AI-Generated Videos from 3D Geometric Temporal Consistency” 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

Wenhan Chen, Sezer Karaoglu, Theo Gevers

Researchers from the University of Amsterdam developed Grab-3D, a detector for AI-generated videos that leverages 3D geometric temporal consistency using vanishing points. It achieves robust cross-domain generalization, outperforming prior methods by over 12% AUC on unseen generative models.

Key insights

  • 3D geometric temporal consistency, especially the stability and smoothness of vanishing point trajectories, provides a robust and generator-agnostic cue for detecting AI-generated videos.
  • AI-generated videos often exhibit 'unstable and jittery' vanishing point trajectories and high-frequency fluctuations in geometric residuals compared to the smooth patterns in real videos.
  • Integrating Geometric Positional Encoding (GPE), Temporal Geometric Attention (TGA), and EMA-based Geometric Residuals into a transformer backbone effectively captures these subtle 3D inconsistencies.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.