CoCoVideo: The High-Quality Commercial-Model-Based Contrastive Benchmark for AI-Generated Video Detection

Authors: Huidong Feng, Wentao Chen, Jie Chen, Xinqi Cai, Ruolong Ma, Yinglin Zheng, Yuxin Lin, Ming Zeng

Published: 2026-05-26 03:18:44+00:00

Comment: Accepected by CVPR 2026

AI Summary

This paper introduces CoCoVideo-26K, a high-quality contrastive benchmark dataset for AI-generated video detection, featuring semantically aligned real-fake video pairs from 13 mainstream commercial generators. To leverage this dataset, they propose CoCoDetect, a novel framework that integrates dual-head contrastive learning with a confidence-gated multimodal large language model (MLLM) inference mechanism. CoCoDetect aims to robustly detect highly realistic video forgeries by analyzing both texture-level artifacts and semantic inconsistencies.

Abstract

With the rapid advancement of artificial intelligence generated content (AIGC) technologies, video forgery has become increasingly prevalent, posing new challenges to public discourse and societal security. Despite remarkable progress in existing deepfake detection methods, AIGC forgery detection remains challenging, as existing datasets mainly rely on open-source video generation models with quality far below that of commercial AIGC systems. Even datasets containing a few commercial samples often retain visible watermarks, compromising authenticity and hindering model generalization to high-fidelity AIGC videos. To address these issues, we introduce CoCoVideo-26K, a contrastive, commercial-model-based AIGC video dataset covering 13 mainstream commercial generators and providing semantically aligned real-fake video pairs. This dataset enables deeper exploration of the differences between authentic and high-quality synthetic videos and establishes a new benchmark for highly realistic video forgery detection. Building on this dataset, we propose CoCoDetect, a detection framework integrating contrastive learning with confidence-gated multimodal large language model (MLLM) inference. An R3D-18 backbone extracts spatio-temporal representations, while a confidence gate routes uncertain cases to an MLLM for reasoning about physical plausibility and scene consistency. Extensive experiments on CoCoVideo-26K and public benchmarks demonstrate state-of-the-art performance, validating the framework's robustness and generalizability. Our code and dataset are available at https://github.com/DonoToT/CoCoVideo.

Key findings

CoCoDetect achieves state-of-the-art performance on the CoCoVideo-26K dataset, outperforming existing methods across various metrics like Accuracy, F1-score, Recall, and AUC. The framework also demonstrates strong cross-dataset generalization capabilities on several public AIGC video benchmarks. Ablation studies confirm that both the contrastive learning component and the confidence-gated MLLM reasoning mechanism are crucial for enhancing detection robustness and generalizability.

Approach

The proposed CoCoDetect framework uses an R3D-18 backbone for spatio-temporal feature extraction, trained with a dual-head design. One head provides authenticity confidence scores via binary cross-entropy, while the other generates contrastive embeddings optimized with a paired contrastive loss. A confidence-gated mechanism routes uncertain predictions to an MLLM, LLaVA-NeXT-Video-7B, for reasoning about physical plausibility and scene consistency, with the MLLM's semantic judgment then adaptively fused with the base model's confidence.

Datasets

CoCoVideo-26K (newly introduced), OpenVid-1M (source for real videos), GVD, GVF, GenVideo, GenVidBench, GenBuster

Model(s)

R3D-18 (backbone), LLaVA-NeXT-Video-7B (Multimodal Large Language Model)

Author countries

China, Singapore

← Previous