Research map / Video / Video segmentation and tracking
AI-Generated Video Detection via Perceptual Straightening
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Detects AI-generated videos by measuring curvature and stepwise distance of DINOv2 frame-embedding trajectories and classifying their aggregated statistics with a lightweight MLP.
Problem
Detecting AI-generated videos generalizes poorly across generators, and existing image- and video-based detectors struggle to capture subtle temporal inconsistencies.
Approach
Inspired by the perceptual straightening hypothesis, natural videos form straighter trajectories in a self-supervised vision transformer's (DINOv2) representation space than AI-generated videos; per-video statistics of stepwise distance and curvature are fed to a lightweight classifier.
Key results
- On VidProM image-detector comparison, ReStraV reaches 97.06% average accuracy and 98.81% mAP.
- On GenVidBench Main task, ReStraV averages 93.01% accuracy; on Plants task it averages 96.96%.
- On Physics-IQ matched pairs, ReStraV attains 97.5% (Pika), 94.9% (Runway), 99.0% (Sora) using mean curvature alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DINOv2 trajectory geometry and an MLP classifier instead of image-based detection | AI-generated video detection · VidProM | accuracy: 63.64 -> 97.06 | Helped |
| Use DINOv2 trajectory geometry and an MLP classifier instead of image-based detection | AI-generated video detection · VidProM | mAP: 69.08 -> 98.81 | Helped |
| Use DINOv2 trajectory geometry for detection | GenVidBench video detectors · GenVidBench Main | accuracy: 79.90 -> 93.01 | Helped |
| Use DINOv2 trajectory geometry for detection | GenVidBench video detectors · GenVidBench Plants | accuracy: 75.09 -> 96.96 | Helped |
| Use higher mean curvature to identify the generated video without classifier training | ReStraV · Physics-IQ matched pairs | 2AFC accuracy: 97.5% for Pika; 94.9% for Runway; 99.0% for Sora | Helped |
| Apply the trained ReStraV MLP to an unseen generator | ReStraV · Veo3 zero-shot | accuracy: 83.2% | Helped |
| Increase analyzed video length from 1 to 5 seconds (ablation) | ReStraV with MLP · VidProM Sora and DVSC2023 | AUROC: exceeding 96% for 2-second segments; approximately 98% for 5-second segments | Helped |
| Sample every 3rd frame (k=3, T=20) instead of sparser or denser sampling (ablation) | ReStraV with MLP · 2-second video sampling | AUROC: approximately 97% | Helped |
| Use sparser sampling (k=5, T=12) rather than k=3 (ablation) | ReStraV with MLP · 2-second video sampling | AUROC | Hurt |
| Use very dense sampling (k=1, T=60) rather than k=3 (ablation) | ReStraV with MLP · 2-second video sampling | AUROC | Hurt |
| Shift the 2-second analysis window across a 5-second video (ablation) | ReStraV with MLP · 5-second videos | detection performance | No clear effect |
| Use DINOv2 instead of pixel-space trajectories | Video trajectory analysis · Natural versus AI-generated videos | curvature separation: Δθ = 45.46° for DINOv2 | Helped |
Surprising findings
- Good perceptual straighteners (HVS-inspired models) straighten AI and natural videos alike, so absolute straightening does not predict detection; DINOv2's differential straightening does.
- Natural videos have lower mean curvature but higher curvature variance than AI-generated videos.
Limitations
- Vulnerable to adversarial training against the detection metric (Goodhart's law concern).
- Future video generators may close the statistical gap between real and generated trajectories.
- Relies on a short 2s window; longer videos may improve performance but are not fully explored.
Experimental setup
Datasets: VidProM DVSC2023 GenVidBench Physics-IQ DeMamba benchmark
Benchmarks: VidProM GenVidBench Physics-IQ DeMamba benchmark Veo3 zero-shot
Metrics: accuracy mAP AUROC F1 latency
Baselines: CNNSpot FreDect Fusing Gram-Net GIA LNP DFD UnivFD TSM X3D MViT V2 SlowFast
Builds on: perceptual straightening hypothesis DINOv2 VideoSwin DeMamba
Closest papers by meaning
- Physics-Driven Spatiotemporal Modeling for AI-Generated Video Detection (NeurIPS 2025 · similarity 0.85)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.84)
- Chirality in Action: Time-Aware Video Representation Learning by Latent Straightening (NeurIPS 2025 · similarity 0.84)
- Manifold Induced Biases for Zero-shot and Few-shot Detection of Generated Images (ICLR 2025 · similarity 0.84)
- Learning predictable and robust neural representations by straightening image sequences (NeurIPS 2024 · similarity 0.84)
- Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.