Deploying AI-generated video detectors in real-world services requires an ultra-low false positive rate (FPR) to avoid mistakenly rejecting authentic content, where standard metrics like AUROC fail to reflect actual operational behavior. We introduce Spatial Patch-Level Incoherence and Temporal Roughness (SPLIT), a training-free detector that operates on patch tokens from a frozen vision encoder to identify both fully generated and partially edited videos.
SPLIT computes two complementary signals:
- Two-step Temporal Roughness (TTR): capturing non-smooth patch trajectories through one-step and two-step feature variation contrast.
- Local Spatial Motion Incoherence (LSMI): measuring spatially inconsistent temporal changes via gradients of a feature-space motion field.
These two signals are fused multiplicatively with gamma correction to enhance real-fake separation at strict thresholds. We further propose a service-aligned evaluation protocol based on Fake Recall at fixed FPR with real-only threshold calibration and cross-real threshold transfer.
Across three benchmarks (FakeParts, GenVideo, and ViF-Bench), SPLIT achieves the highest Fake Recall at FPR = $0.1\%$, significantly outperforming both supervised and training-free baselines while remaining robust to post-processing with negligible overhead. The code is publicly available at GitHub.
Blogger's Review: The introduction of SPLIT offers a fresh perspective in the realm of video detection, particularly in its efficiency and accuracy when handling generated and edited content. Its innovative signal fusion mechanism significantly enhances detection performance, making it a promising candidate for widespread practical application.