SPLIT: Training-Free AI-Generated and Partially Edited Video Detection via Spatial Patch-Level Incoherence and Temporal Roughness

2026-07-02

AI Summary

Deploying AI-generated video detectors in real-world scenarios demands effective identification of fully synthetic or partially edited videos under extremely low false positive rates (e.g., FPR = 0.1%), a requirement poorly captured by conventional metrics such as AUROC. To address this, this work proposes a training-free detection method that leverages patch-level features from a frozen visual encoder, introducing for the first time a combination of two-step temporal roughness (TTR) and local spatial motion inconsistency (LSMI). Multiplicative fusion and gamma correction are employed to enhance the separation between real and fake content. The approach significantly outperforms both supervised and training-free baselines on FakeParts, GenVideo, and ViF-Bench, demonstrating strong robustness to post-processing and minimal computational overhead, while also introducing an evaluation protocol tailored for practical deployment.

Abstract

Deploying AI-generated video detectors in real-world services demands an ultra-low false positive rate (FPR) on real videos to avoid falsely rejecting authentic content, a regime where standard metrics such as AUROC fail to reflect actual operating 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 detect both fully generated and partially edited videos. SPLIT computes two complementary signals: Two-step Temporal Roughness (TTR), capturing non-smooth patch trajectories via one-step and two-step feature variation contrast, and Local Spatial Motion Incoherence (LSMI), measuring spatially inconsistent temporal changes through gradients of a feature-space motion field. The two are fused multiplicatively with gamma correction to sharpen 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 = $0.1\%$, substantially outperforming supervised and training-free baselines while remaining robust to post-processing with negligible overhead. The code is publicly available at https://github.com/mldljyh/SPLIT .

Problem

• AI-generated video detection
• false positive rate
• real-world deployment
• partially edited video
• evaluation protocol

Innovation

• training-free detection
• temporal roughness
• spatial incoherence
• low false positive rate
• AI-generated video detection