Title: CoCoVideo: A Contrastive Benchmark for AI-Generated Video Detection Based on High-Quality Commercial Models

Abstract

The proliferation of artificial intelligence-generated content (AIGC) has intensified the prevalence of video forgery, creating significant threats to public discourse and societal security. While current deepfake detection techniques have advanced, identifying AIGC forgeries remains difficult because most existing datasets are built upon open-source video generation models that produce lower-quality output compared to commercial AIGC systems. Furthermore, datasets that do include some commercial samples often contain visible watermarks, which undermines authenticity and limits the ability of models to generalize to high-fidelity synthetic videos.

To overcome these limitations, we present CoCoVideo-26K, a novel contrastive dataset based on commercial models. This resource covers 13 leading commercial generators and provides semantically aligned pairs of real and fake videos. CoCoVideo-26K facilitates a deeper analysis of the distinctions between authentic footage and high-quality synthetic media, setting a new standard for detecting highly realistic video forgeries.

Leveraging this dataset, we introduce CoCoDetect, a detection framework that combines contrastive learning with confidence-gated multimodal large language model (MLLM) inference. The system utilizes an R3D-18 backbone to extract spatio-temporal features. A confidence gate mechanism directs uncertain cases to an MLLM, which performs reasoning regarding scene consistency and physical plausibility. Comprehensive experiments conducted on CoCoVideo-26K and various public benchmarks confirm that our approach achieves state-of-the-art performance, demonstrating both robustness and strong generalizability. The dataset and source code are accessible at https://github.com/DonoToT/CoCoVideo.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC