AI-Generated Video Detection
Perceptual Straightening: A Novel Approach to Distinguish AI-Generated Videos
The rapid advancement of generative AI enables highly realistic synthetic videos, posing significant challenges for content authentication and raising urgent concerns about misuse. Existing detection methods often struggle with generalization and capturing subtle temporal inconsistencies. We propose ReStraV (Representation Straightening for Video), a novel approach to distinguish natural from AI-generated videos. Inspired by the "perceptual straightening" hypothesis [1, 2]—which suggests real-world video trajectories become more straight in neural representation domain—we analyze deviations from this expected geometric property. Using a pre-trained self-supervised vision transformer (DINOv2), we quantify the temporal curvature and stepwise distance in the model's representation domain. We aggregate statistics of these measures for each video and train a classifier. Our analysis shows that AI-generated videos exhibit significantly different curvature and distance patterns compared to real videos. A lightweight classifier achieves state-of-the-art detection performance (e.g., 97.17% accuracy and 98.63% AUROC on the VidProM benchmark [3]), substantially outperforming existing image- and video-based methods. ReStraV is computationally efficient, offering a low-cost and effective detection solution. This work provides new insights into using neural representation geometry for AI-generated video detection.
Authors: Christian Internò, Robert Geirhos, Barbara Hammer, Markus Olhofer, Sunny Liu, David Klindt
Executive Impact: Key Performance Indicators
ReStraV sets a new standard for AI-generated video detection, offering unparalleled accuracy and efficiency crucial for maintaining digital trust.
0
Detection Accuracy
0
AUROC Score
0
Inference Time (per video)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
ReStraV: Leveraging Perceptual Straightening
ReStraV introduces a novel approach to AI-generated video detection by leveraging the "perceptual straightening" hypothesis. This principle posits that natural video trajectories appear straighter in neural representation space compared to synthetic ones. By analyzing deviations from this expected geometric property using DINOv2, ReStraV effectively identifies AI-generated content.
Enterprise Process Flow
Frames processed by DINOv2 SSE
→
Embeddings Collected
→
Natural Videos Trace Straighter Paths in Representation Space
→
Trajectory Geometry Provides Discriminative Signal
→
Classifier: AI-generated vs. Natural
| Feature | Pixel Space Behavior | DINOv2 Representation Space Behavior |
|---|---|---|
| Curvature | Substantial overlap between natural and AI-generated video trajectories. | Natural trajectories straighten significantly, leading to clear separation from AI-generated videos. |
| Distance (Stepwise) | Similar overlap between natural and AI-generated video trajectories. | Natural videos exhibit distinct and consistent stepwise distance patterns compared to AI-generated. |
State-of-the-Art Detection Performance
ReStraV consistently outperforms existing methods across various benchmarks, demonstrating robust and reliable detection capabilities for AI-generated video content.
97.06%
Average Accuracy on VidProM Benchmark
98.81%
Average mAP (Mean Average Precision) on VidProM
| Method | Average Accuracy (%) | Average mAP (%) |
|---|---|---|
| Gram-Net | 63.64 | 69.08 |
| FreDect | 60.98 | 60.99 |
| LNP | 45.32 | 43.70 |
| ReStraV | 97.06 | 98.81 |
Robust Generalization Across Diverse Scenarios
ReStraV demonstrates strong performance even against unseen and future generative models, and on challenging datasets like 'Plants', highlighting its broad applicability and resilience.
| Condition | VideoSwin Acc (%) | ReStraV Acc (%) | VideoSwin mAP (%) | ReStraV mAP (%) |
|---|---|---|---|---|
| Seen generators | 77.91 | 97.05 | 75.33 | 98.78 |
| Unseen generators | 62.44 | 89.45 | 59.61 | 97.32 |
| Future generators (Sora) | 60.70 | 80.05 | 58.20 | 92.85 |
| Method | Avg Acc (Main) (%) | Avg Acc (Plants) (%) |
|---|---|---|
| TSM | 76.40 | 55.30 |
| X3D | 77.09 | 62.15 |
| MVIT V2 | 79.90 | 52.86 |
| ReStraV | 93.01 | 96.96 |
Robustness to Scene Cuts
An analysis of 13,000 AI-generated and 13,000 natural videos revealed that the average number of hard scene cuts is low and comparable across both categories. This confirms that ReStraV's high performance is not confounded by scene cut frequency, but rather by subtle geometric inconsistencies in AI-generated video trajectories.
Broader Implications and Future Outlook
ReStraV offers profound insights into both AI detection mechanisms and the underlying computational principles of perception, while also acknowledging the dynamic challenges of the AI landscape.
Implications for Neuroscience
The finding that natural videos trace straighter paths than AI-generated ones in a frozen vision transformer dovetails with the perceptual straightening phenomenon reported in perceptual decision tasks and brain recordings. This suggests a shared computational pressure, across biological and artificial systems, to encode 'intuitive physics' in a geometry that favors smooth temporal trajectories, opening new avenues for understanding brain function.
Addressing Limitations and the 'Arms Race'
The authors acknowledge the 'Goodhart's law' concern, where detection measures could become targets for generative models. Future work includes exploring multi-metric detection and advancing complementary methods like robust content watermarking. ReStraV also highlights the ongoing 'arms race' in AI detection and the need for continuous research and robust safety measures to combat misuse and ensure digital trust.
Calculate Your Potential AI Impact
Estimate the transformative efficiency gains and cost savings AI can bring to your enterprise operations.
Annual Savings Estimate
$0
Annual Hours Reclaimed
0
Your Enterprise AI Implementation Roadmap
A structured approach to integrating cutting-edge AI, ensuring seamless adoption and measurable results.
Phase 1: Discovery & Strategy
Initial consultation and deep dive into your current operations, pain points, and strategic objectives. We identify key opportunities for AI integration and define success metrics tailored to your business.
Phase 2: Pilot Program & Validation
Develop and deploy a targeted AI pilot program within a controlled environment. This phase focuses on validating the AI model's effectiveness, gathering feedback, and fine-tuning parameters for optimal performance.
Phase 3: Scaled Deployment & Integration
Full-scale integration of the AI solution across your enterprise, meticulously ensuring compatibility with existing systems and workflows. Comprehensive training and support are provided to your teams.
Phase 4: Optimization & Future-Proofing
Continuous monitoring, performance optimization, and regular updates to adapt to evolving business needs and technological advancements. We ensure your AI remains a competitive asset.
Ready to Transform Your Enterprise with AI?
Book a personalized consultation with our AI experts to explore how these advanced strategies can drive innovation and efficiency in your organization.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat