DAVID-XR1: Detecting AI-Generated Videos with Explainable Reasoning

TL;DR

DAVID-XR1 integrates fine-grained defect annotations and chain-of-thought reasoning, enabling interpretable AI-generated video detection with strong cross-generator generalization.

cs.CV πŸ”΄ Advanced 2025-06-13 8 citations 43 views
Yifeng Gao Yifan Ding Hongyu Su Juncheng Li Yunhan Zhao Lin Luo Zixing Chen Li Wang Xin Wang Yixu Wang Xingjun Ma Yu-Gang Jiang
AI Video Detection Explainability Multimodal Models Chain-of-Thought Defect Annotation

Key Findings

Methodology

This paper introduces DAVID-XR1, a multimodal model built upon the richly annotated dataset DAVID-X. The dataset includes defect categories, spatio-temporal localization, and natural language rationales, mimicking human reasoning. The model employs a visual Chain-of-Thought (CoT) framework, distilled from Gemini 2.5 Pro, to generate an explicit reasoning chain that identifies and explains artifacts indicative of AI generation. During training, a joint optimization strategy combines the language modeling objective with a binary classifier that predicts real vs. AI-generated videos based on the reasoning path. The model leverages chain-of-thought distillation to improve interpretability and generalization, achieving high accuracy across diverse generation modes and unseen generators.

Key Results

  • The model achieves 76.7% detection accuracy on out-of-domain datasets, including unseen generators like Veo2 and Pika v2.2, outperforming baseline black-box classifiers by over 40%. Explanation precision reaches 54.7%, demonstrating the model's ability to generate human-readable rationales that specify defect types, locations, and temporal spans. Ablation studies confirm that defect categorization, spatio-temporal localization, and natural language explanations synergistically enhance detection robustness, especially in challenging scenarios.
  • Cross-generator tests show the model maintains high performance, with detection accuracy above 70% on unseen models, indicating excellent generalization. The joint training with chain-of-thought pathways significantly improves both detection and interpretability compared to language-only fine-tuning. The model effectively localizes subtle artifacts such as object inconsistency, texture jitter, and lighting anomalies, providing detailed evidence for human auditors.
  • The experiments validate that integrating rich defect annotations and explicit reasoning paths leads to more trustworthy detection systems. The model's ability to produce step-by-step explanations aligns with the increasing demand for transparency in AI decision-making, especially in sensitive applications like digital forensics and content moderation.

Significance

This work marks a significant advancement in AI-generated video detection by embedding interpretability into the core of the model. Moving beyond black-box classifiers, DAVID-XR1 offers transparent reasoning, fostering trust and facilitating human verification. Its strong cross-generator generalization addresses a critical challenge in deepfake detection, making it suitable for real-world deployment in social media moderation, legal investigations, and misinformation control. The dataset DAVID-X further provides a valuable resource for future research in explainable multimodal forensics, setting a new standard for dataset quality and annotation granularity.

Technical Contribution

The paper introduces a novel combination of fine-grained defect annotations, spatio-temporal localization, and chain-of-thought reasoning within a multimodal framework. The use of Gemini 2.5 Pro for knowledge distillation enables the extraction of explicit reasoning paths, which are incorporated into a supervised fine-tuning process that jointly optimizes detection and explanation. The model architecture extends existing vision-language models by integrating a lightweight binary classifier that aligns reasoning paths with final decisions, ensuring interpretability. The dataset DAVID-X provides detailed defect-level annotations, including defect categories, spatial points, and temporal spans, serving as high-quality supervision for training and evaluation.

Novelty

This research is pioneering in combining defect-level annotations with visual chain-of-thought reasoning for AI-generated video detection. Unlike prior methods that treat detection as a black-box classification, this approach produces transparent, human-readable explanations grounded in fine-grained evidence. The integration of Gemini 2.5 Pro's reasoning paths into a multimodal detection framework, coupled with joint optimization, offers a new paradigm for trustworthy deepfake detection. The model's ability to generalize across unseen generators and modes further distinguishes it from existing solutions.

Limitations

  • Despite high accuracy, the model's performance may degrade on videos with extremely subtle or heavily obscured defects, especially in very low-quality or highly compressed videos.
  • The reliance on detailed annotations increases data collection and labeling costs, which may hinder large-scale deployment.
  • The computational complexity of chain-of-thought reasoning and pixel-level localization limits real-time application, requiring further optimization for deployment in live scenarios.
  • The model's robustness against future, more sophisticated generative models remains to be tested, necessitating continual dataset updates and model retraining.

Future Work

Future research will focus on reducing inference latency through model compression and architecture optimization, enabling real-time detection. Expanding the dataset to include more diverse generation techniques and real-world scenarios will enhance robustness. Exploring semi-supervised learning and active annotation strategies can reduce labeling costs. Additionally, integrating audio and text modalities may further improve detection accuracy and explanation richness. Developing adaptive models that can quickly incorporate new generative techniques will be crucial for maintaining effectiveness against evolving deepfake technologies.

AI Executive Summary

The rapid proliferation of AI-generated videos has transformed digital media, enabling creators to produce highly realistic content with minimal effort. Technologies like Veo 2, Kling, and Pika have made it possible to generate videos that are nearly indistinguishable from authentic footage, raising serious concerns about misinformation, fraud, and privacy. Existing detection methods primarily rely on binary classifiers trained on large datasets, which, while effective in some scenarios, lack transparency and often fail to generalize to unseen generation techniques.

Recognizing the critical need for trustworthy detection, this research introduces DAVID-XR1, a novel multimodal model that combines detailed defect annotations with chain-of-thought reasoning to produce transparent, explainable decisions. The core innovation lies in constructing a richly annotated dataset, DAVID-X, which includes defect categories, spatio-temporal localization, and human-readable rationales. Leveraging this dataset, the authors distill reasoning paths from Gemini 2.5 Pro, transforming complex inference into explicit, step-by-step explanations that identify artifacts such as object inconsistency, texture jitter, and lighting anomalies.

The model, DAVID-XR1, employs a visual chain-of-thought framework, jointly optimizing detection accuracy and explanation quality. Experimental results demonstrate that it achieves 76.7% accuracy on out-of-domain datasets, significantly outperforming traditional black-box classifiers. Its explanations, with a precision of 54.7%, provide detailed evidence that can be verified by human auditors, fostering greater trust in automated detection systems.

This work has profound implications for digital media integrity, content moderation, and forensic investigations. By making the detection process transparent and grounded in concrete evidence, it addresses the core challenge of establishing trustworthiness in AI-generated content. Future directions include enhancing real-time performance, expanding dataset diversity, and integrating additional modalities to further strengthen detection robustness. Overall, DAVID-XR1 sets a new standard for explainable, generalizable AI-generated video detection, paving the way for safer and more trustworthy digital environments.

Deep Dive

Abstract

As AI-generated video becomes increasingly pervasive across media platforms, the ability to reliably distinguish synthetic content from authentic footage has become both urgent and essential. Existing approaches have primarily treated this challenge as a binary classification task, offering limited insight into where or why a model identifies a video as AI-generated. However, the core challenge extends beyond simply detecting subtle artifacts; it requires providing fine-grained, persuasive evidence that can convince auditors and end-users alike. To address this critical gap, we introduce DAVID-X, the first dataset to pair AI-generated videos with detailed defect-level, temporal-spatial annotations and written rationales. Leveraging these rich annotations, we present DAVID-XR1, a video-language model designed to deliver an interpretable chain of visual reasoning-including defect categorization, temporal-spatial localization, and natural language explanations. This approach fundamentally transforms AI-generated video detection from an opaque black-box decision into a transparent and verifiable diagnostic process. We demonstrate that a general-purpose backbone, fine-tuned on our compact dataset and enhanced with chain-of-thought distillation, achieves strong generalization across a variety of generators and generation modes. Our results highlight the promise of explainable detection methods for trustworthy identification of AI-generated video content.

cs.CV cs.AI

References (20)

VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models

Haoxin Chen, Yong Zhang, Xiaodong Cun et al.

2024 682 citations ⭐ Influential View Analysis β†’

Open-Sora: Democratizing Efficient Video Production for All

Zangwei Zheng, Xiangyu Peng, Tianji Yang et al.

2024 785 citations ⭐ Influential View Analysis β†’

HunyuanVideo: A Systematic Framework For Large Video Generative Models

Weijie Kong, Qi Tian, Zijian Zhang et al.

2024 1613 citations ⭐ Influential View Analysis β†’

Open-Sora Plan: Open-Source Large Video Generation Model

Bin Lin, Yunyang Ge, Xinhua Cheng et al.

2024 310 citations ⭐ Influential View Analysis β†’

ModelScope Text-to-Video Technical Report

Jiuniu Wang, Hangjie Yuan, Dayou Chen et al.

2023 736 citations ⭐ Influential View Analysis β†’

CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

Zhuoyi Yang, Jiayan Teng, Wendi Zheng et al.

2024 2305 citations ⭐ Influential View Analysis β†’

Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets

A. Blattmann, Tim Dockhorn, Sumith Kulal et al.

2023 2855 citations View Analysis β†’

OpenAI o1 System Card

Ahmed El-Kishky

2024 2012 citations

Gradio: Hassle-Free Sharing and Testing of ML Models in the Wild

Abubakar Abid, Ali Abdalla, Ali Abid et al.

2019 367 citations View Analysis β†’

K-Means Clustering

Xin Jin, Jiawei Han

2004 732 citations

UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

Leland McInnes, John Healy

2018 13180 citations View Analysis β†’

AI-Generated Video Detection via Spatial-Temporal Anomaly Learning

Jianfa Bai, Man Lin, Gang Cao et al.

2024 43 citations

VideoCrafter1: Open Diffusion Models for High-Quality Video Generation

Haoxin Chen, Menghan Xia, Yin-Yin He et al.

2023 613 citations View Analysis β†’

FETV: A Benchmark for Fine-Grained Evaluation of Open-Domain Text-to-Video Generation

Yuanxin Liu, Lei Li, Shuhuai Ren et al.

2023 143 citations View Analysis β†’

Latent Video Diffusion Models for High-Fidelity Long Video Generation

Yin-Yin He, Tianyu Yang, Yong Zhang et al.

2022 442 citations View Analysis β†’

Compositional Chain-of-Thought Prompting for Large Multimodal Models

Chancharik Mitra, Brandon Huang, Trevor Darrell et al.

2023 235 citations View Analysis β†’

VBench: Comprehensive Benchmark Suite for Video Generative Models

Ziqi Huang, Yinan He, Jiashuo Yu et al.

2023 1838 citations View Analysis β†’

V*: Guided Visual Search as a Core Mechanism in Multimodal LLMs

Penghao Wu, Saining Xie

2023 590 citations View Analysis β†’

Intern VL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks

Zhe Chen, Jiannan Wu, Wenhai Wang et al.

2023 3299 citations View Analysis β†’

Detecting AI-Generated Video via Frame Consistency

Long Ma, Zhiyuan Yan, Qinglang Guo et al.

2024 23 citations View Analysis β†’

Cited By (8)

VidForensics-M1: Meta-Detection Reinforcement Learning with Verifiable Temporal Grounding for AI-Generated Video Forensics

2026 ⭐ Influential View Analysis β†’

VideoVeritas: AI-Generated Video Detection via Perception Pretext Reinforcement Learning

2026 12 citations ⭐ Influential View Analysis β†’

Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning

2025 19 citations ⭐ Influential View Analysis β†’

SafeGuard: A Multi-Agent Perception-Reasoning Framework for Social-Risk AI-Generated Video Detection

Explainable Forensics of Manipulated Segments in Untrimmed Long Videos

GenVideoLens: Where LVLMs Fall Short in AI-Generated Video Detection?

2026 1 citations View Analysis β†’

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

2026 4 citations View Analysis β†’

Training-free Detection of Text-to-video Generations via Over-coherence

2026 4 citations