Your One-Stop Solution for AI-Generated Video Detection

TL;DR

This work tackles the problem of reliable AI-generated video detection by introducing AIGVDBench, a large-scale, standardized benchmark that spans $31$ generation methods and over $440{,}000$ videos, enabling over $1{,}500$ evaluations on $33$ detectors across four task categories. It addresses dual bottlenecks in prior work—dataset scale/diversity and benchmark-driven analysis—via a principled pipeline: attribute-balanced prompt selection, careful open/closed-source model curation, and compression unification to $H.264$. Through extensive evaluations and eight in-depth analyses, the authors demonstrate that higher generative quality does not simplify detection, reveal nuanced cross-task performance patterns, and show Vision-Language Models are currently limited for this task, while traditional video and image detectors with tailored training/data remain most effective. The benchmark’s findings offer actionable directions for future detector design, data curation, and the exploration of interpretable, cross-modal detection approaches, ultimately contributing to more trustworthy digital media authentication. The work thus provides a solid foundation and a practical toolset for accelerating research in AI-generated video detection.

Abstract

Recent advances in generative modeling can create remarkably realistic synthetic videos, making it increasingly difficult for humans to distinguish them from real ones and necessitating reliable detection methods. However, two key limitations hinder the development of this field. \textbf{From the dataset perspective}, existing datasets are often limited in scale and constructed using outdated or narrowly scoped generative models, making it difficult to capture the diversity and rapid evolution of modern generative techniques. Moreover, the dataset construction process frequently prioritizes quantity over quality, neglecting essential aspects such as semantic diversity, scenario coverage, and technological representativeness. \textbf{From the benchmark perspective}, current benchmarks largely remain at the stage of dataset creation, leaving many fundamental issues and in-depth analysis yet to be systematically explored. Addressing this gap, we propose AIGVDBench, a benchmark designed to be comprehensive and representative, covering \textbf{31} state-of-the-art generation models and over \textbf{440,000} videos. By executing more than \textbf{1,500} evaluations on \textbf{33} existing detectors belonging to four distinct categories. This work presents \textbf{8 in-depth analyses} from multiple perspectives and identifies \textbf{4 novel findings} that offer valuable insights for future research. We hope this work provides a solid foundation for advancing the field of AI-generated video detection. Our benchmark is open-sourced at https://github.com/LongMa-2025/AIGVDBench.

Figures (30)

• Figure 1: Overview of AIGVDBench. Our benchmark demonstrates advantages in scale, content diversity, coverage of generation tasks, diversity of video generation models, and video realism.
• Figure 2: Visual examples of the AIGVDBench.
• Figure 3: Each heatmap displays the AUC scores of a detector trained on one generative model (rows) and evaluated across others (columns).
• Figure 4: Comparison of video generation models on Aesthetic Quality, Background Consistency, Dynamic Degree, Imaging Quality, Motion Smoothness, Subject Consistency, Frame-Level (Image & Aesthetic Average) and Final Score based on Vbench. vbench
• Figure 5: AUC Performance Comparison of Video Classification, Generated Image Detection, and Generated Video Detection Models across Closed-Source Generation Models.