Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods

2026-03-16

AI Summary

This work addresses the growing threat of synthetic videos being exploited for disinformation, a challenge exacerbated by limitations in current detection approaches that either overlook temporal cues or rely on supervised training, thereby compromising generalization. To overcome these issues, we propose STALL—a training-free, model-agnostic, zero-shot method for video authenticity verification. STALL leverages a spatio-temporal joint probability model grounded in the statistical properties of real-world data, integrating spatial and temporal evidence through a theoretically justified likelihood scoring mechanism to effectively discriminate generated videos. Evaluated on two established benchmarks and a newly introduced dataset, ComGenVid, STALL substantially outperforms existing image- and video-based detection baselines, marking the first demonstration of efficient and generalizable zero-shot detection of AI-generated videos.

Technology Category

• Computer Vision: Video Understanding & Activity Analysis
• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Web Mining and Content Analysis: Web data provenance, reliability, and authenticity

Abstract

Following major advances in text and image generation, the video domain has surged, producing highly realistic and controllable sequences. Along with this progress, these models also raise serious concerns about misinformation, making reliable detection of synthetic videos increasingly crucial. Image-based detectors are fundamentally limited because they operate per frame and ignore temporal dynamics, while supervised video detectors generalize poorly to unseen generators, a critical drawback given the rapid emergence of new models. These challenges motivate zero-shot approaches, which avoid synthetic data and instead score content against real-data statistics, enabling training-free, model-agnostic detection. We introduce \emph{STALL}, a simple, training-free, theoretically justified detector that provides likelihood-based scoring for videos, jointly modeling spatial and temporal evidence within a probabilistic framework. We evaluate STALL on two public benchmarks and introduce ComGenVid, a new benchmark with state-of-the-art generative models. STALL consistently outperforms prior image- and video-based baselines. Code and data are available at https://omerbenhayun.github.io/stall-video.

Problem

• generated video detection
• zero-shot detection
• temporal dynamics
• model-agnostic
• misinformation

Innovation

• training-free
• zero-shot detection
• spatial-temporal likelihood
• model-agnostic
• deepfake detection