G2VD: A New Framework for Generalizable AI-Generated Video Detection

Researchers have proposed G2VD, a generalizable AI-generated video detection framework that addresses the problem of shortcut learning in existing detectors. Current methods perform well on known generators but degrade on unseen ones due to reliance on domain-specific bias. G2VD uses counterfactual intervention and causal disentanglement to improve cross-domain generalization. The framework includes a counterfactual intervention pipeline (CFIPipeline) that constructs counterfactual samples via VAE-based reconstruction and frequency-domain and pixel-domain alignment, weakening spurious correlations. A causal disentanglement classifier then leverages these samples for robust detection. The paper is available on arXiv under ID 2607.04607.