Research map / Video / Video segmentation and tracking
Physics-Driven Spatiotemporal Modeling for AI-Generated Video Detection
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.
Proposes NSG-VD, a physics-driven AI-generated video detector that uses a diffusion-based estimate of a probability-flow "Normalized Spatiotemporal Gradient" statistic and MMD against real-video references.
Problem
Detecting AI-generated videos (including highly realistic ones like Sora) when subtle spatiotemporal inconsistencies and physics-law violations are hard to capture with artifact-based methods.
Approach
Derives a Normalized Spatiotemporal Gradient (NSG) from the continuity equation of probability flow, estimates it with pre-trained diffusion score functions plus a brightness-constancy temporal approximation, and detects fakes via deep-kernel MMD between NSG features of test and real videos.
Key results
- NSG-VD reaches 88.02% average Recall vs 72.02% for DeMamba when trained on Kinetics-400 and Pika, a 16.00% gain.
- Trained on Kinetics-400 and SEINE, NSG-VD achieves 97.13% average Recall and 87.45% F1 vs DeMamba's 71.25% Recall.
- In the data-imbalanced setting (1,000 SEINE fakes), NSG-VD reaches 82.14% Recall on Sora vs at most 33.93% for baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NSG-VD instead of DeMamba | GenVideo detection trained with Kinetics-400 and Pika · GenVideo test, average across generators | Recall: 16.00% ↑ | Helped |
| use NSG-VD instead of DeMamba | GenVideo detection trained with Kinetics-400 and SEINE · GenVideo test, average across generators | Recall: 25.88% ↑ | Helped |
| use NSG-VD instead of DeMamba | GenVideo detection trained with 1,000 SEINE generated videos · GenVideo test, average across generators | Recall: 29.12% ↑ | Helped |
| combine spatial gradients with temporal derivatives in NSG (ablation) | NSG-VD · GenVideo test, Kinetics-400 + Pika training, average | F1: 83.40 -> 90.87 | Helped |
| add temporal derivatives to spatial gradients (ablation) | NSG-VD · GenVideo test, Kinetics-400 + Pika training, average | Recall: 87.99 -> 88.02 | Helped |
| add spatial gradients to temporal derivatives (ablation) | NSG-VD · GenVideo test, Kinetics-400 + Pika training, average | F1: 66.97 -> 90.87 | Helped |
| vary the decision threshold across τ ∈ [0.7, 1.1] (ablation) | NSG-VD · GenVideo test across diverse generators | Recall | No clear effect |
Surprising findings
- Temporal derivative component alone performs weakly (60.35% Recall), while spatial gradients alone reach 87.99% Recall.
Limitations
- Theoretical analysis relies on Gaussian assumptions for video distributions, which may not hold for real video.
- Assumes incompressible flow approximation (divergence term negligible) and brightness constancy, which are approximations.
Experimental setup
Datasets: GenVideo Kinetics-400 SEINE Pika MSR-VTT
Benchmarks: GenVideo
Metrics: Recall Accuracy F1-score AUROC
Baselines: DeMamba NPR TALL STIL
Builds on: Maximum Mean Discrepancy deep kernel learning brightness constancy assumption score-based diffusion models
Closest papers by meaning
- Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection (NeurIPS 2026 · similarity 0.87)
- GEOPHYS: The Geometry of Physical Plausibility (NeurIPS 2026 · similarity 0.86)
- Video Anomaly Detection with Motion and Appearance Guided Patch Diffusion Model (AAAI 2025 · similarity 0.85)
- AI-Generated Video Detection via Perceptual Straightening (NeurIPS 2025 · similarity 0.85)
- FlowMo: Variance-Based Flow Guidance for Coherent Motion in Video Generation (NeurIPS 2025 · similarity 0.85)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.85)
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.