Video Forensic Self-Descriptions: Leveraging Temporally Distributed Forensic Microstructures for Zero-Shot Detection of AI-Generated Videos

Azizpour, Aref; Nguyen, Tai D.; Stamm, Matthew C.

Video Forensic Self-Descriptions: Leveraging Temporally Distributed Forensic Microstructures for Zero-Shot Detection of AI-Generated Videos

Aref Azizpour¹, Tai D. Nguyen¹, Matthew C. Stamm¹

¹Department of Electrical and Computer Engineering, Drexel University

🏆 NeurIPS 2026 (Oral Presentation)

Code Project Page Paper (Coming Soon) arXiv (Coming Soon)

Figure 1: Overview of the Video Forensic Self-Description (VFSD) framework. We model a video scene as a dynamic system and extract temporally distributed forensic microstructures. Two linear convolutional predictors (visual evolution f_I and motion evolution f_O) are trained exclusively on real video to model visual and motion evolution. 3D spatio-temporal autoregressive (AR) models fit the resulting prediction residuals to produce compact self-descriptions (VE-FSD Φ and ME-FSD Z) that naturally cluster real videos, enabling zero-shot detection via Gaussian Mixture Models (GMMs) and high-accuracy source generator attribution across 35 generative models.

Abstract

AI-generated videos are becoming increasingly realistic, yet most detection methods focus on spatial artifacts within individual frames and require synthetic training data from known generators, causing them to fail on unseen generators. We identify two previously unexplored temporally distributed forensic microstructures that reveal whether a video was AI-generated. These traces arise from how a video's visual field and motion evolve over time, capturing subtle deviations that generators introduce when approximating real-world temporal dynamics.

To expose them, we train two simple predictors on real videos only, modeling visual evolution (VE-FSD) and motion evolution (ME-FSD), and analyze their prediction residuals. We then fit 3D spatio-temporal autoregressive models to these residuals, producing compact descriptors called Video Forensic Self-Descriptions (VFSD). VFSDs of real videos naturally cluster together and separate from AI-generated ones, enabling zero-shot detection without synthetic training data and source attribution. Experiments across multiple datasets and 35 generators demonstrate VFSD achieves state-of-the-art performance on both tasks.

Key Contributions

1. Temporal Microstructures

Identifies a novel class of forensic evidence: temporally distributed microstructures that emerge from deviations in how visual appearance and motion dynamics evolve across successive frames, complementing spatial single-frame artifacts.

2. Real-Only Predictors

Trains decoupled linear convolutional predictors (f_I, f_O) solely on real video to learn genuine scene physics. Once trained, they remain permanently frozen, enabling generalizable zero-shot detection without ever seeing synthetic video.

3. 3D AR Self-Descriptions

Parameterizes spatial and cross-frame residual correlations into compact 3D autoregressive (AR) fingerprints. Each video describes its own microstructural dynamics, supporting both zero-shot outlier scoring and multi-class source attribution.

The Physical Intuition: Artificial Smoothness

Why are temporal prediction residuals so discriminative between real and AI-generated video?

Current video generators enforce temporal consistency using 3D convolutions and cross-frame attention mechanisms. Because these models struggle to replicate the complex, non-linear stochastic variations of real-world physics, they produce artificially smoothed, low-entropy motion trajectories.

A counterintuitive consequence is that our motion predictor (f_O) achieves lower prediction error on synthetic videos than on real videos. As shown in the KDE plot, motion residuals of real videos exhibit substantially higher variance and energy, while AI-generated videos form tighter, artificially low-energy distributions.

Figure 2: Motion residual energy distributions for real vs. AI-generated video generators.

Visualizing Forensic Microstructures

By computing the 2D FFT power spectrum of the fitted 3D autoregressive self-descriptions, distinct frequency response patterns emerge across real videos and different generative pipelines.

Stream	Real MIT	Gen Runway	Gen Luma AI	Gen CogVidX	Gen Midjourney	Gen Sora	Gen Mochi
ME-FSD (Motion)
VE-FSD (Visual)

Figure 3: AR coefficient frequency response spectra. Real video residuals produce broad, smoothly varying low-frequency structures reflecting natural sensor noise and model mismatch. In stark contrast, each generative model introduces distinct, higher-frequency artifacts and directional ringing that act as identifiable architectural fingerprints.

Experimental Benchmarks

1. Zero-Shot AI Video Detection

We evaluate 105 real-vs-synthetic test combinations spanning 35 generative video models across three real benchmarks (Moments-in-Time, Video-ACID, and GenVideo). Competing methods utilize officially released pre-trained weights, whereas VFSD operates strictly zero-shot without observing any synthetic video during training.

Method	Type	GenVideo (AUC)	Video-ACID (AUC)	MIT (AUC)	Average (AUC)
NPR (CVPR 2024)	Image-Based	0.839	0.548	0.730	0.706
RINE (ECCV 2024)	Image-Based	0.744	0.825	0.879	0.816
FSD (CVPR 2025)	Image-Based	0.999	0.755	0.737	0.830
DeMamba (arXiv 2024)	Video-Based	0.451	0.568	0.534	0.527
D3 (ICCV 2025)	Video-Based	0.928	0.844	0.688	0.820
AIGVdet (PRCV 2024)	Video-Based	0.875	0.777	0.805	0.819
FTCN (ICCV 2021)	Video-Based	0.579	0.596	0.548	0.547
Vahdati et al. (CVPRW 2024)	Video-Based	0.868	0.770	0.827	0.822
ME-FSD (Ours)	Temporal Motion	0.998	0.932	0.750	0.893
VE-FSD (Ours)	Temporal Visual	1.000	0.930	0.831	0.920
VFSD (Ours Combined)	Dual Stream	1.000	0.957	0.791	0.916

2. Source Generator Attribution & t-SNE Clustering

Beyond binary detection, we test whether self-descriptions capture generator-specific fingerprints. While general vision models (CLIP, DINOv2) and spatial detectors (NPR, FSD) suffer heavy source entanglement under KNN classification, VFSD achieves 0.959 accuracy with linear SVM and 0.851 with KNN across 7 diverse sources, proving that temporal microstructures are inherently source-discriminative.

Figure 4: 2D t-SNE projections comparing baseline representations vs. VFSD. VFSD cleanly maps each generative source into distinct, well-separated clusters while cleanly isolating real videos.

Embedding Method	KNN Acc.	SVM Acc.
VISION EMBEDDINGS
ResNet-50	0.49	0.63
CLIP-ViT	0.65	0.81
DINOv2	0.54	0.63
FORENSIC EMBEDDINGS
NPR (CVPR 2024)	0.39	0.83
D3 (ICCV 2025)	0.65	0.82
Vahdati et al. (CVPRW 2024)	0.37	0.48
FSD (CVPR 2025)	0.61	0.77
ME-FSD (Ours)	0.83	0.95
VE-FSD (Ours)	0.78	0.92
VFSD (Ours Combined)	0.85	0.96

Notice how KNN accuracy for baselines collapses (e.g. NPR drops from 0.830 to 0.390), indicating their features are not naturally clustered, whereas VFSD retains high accuracy (0.851) even with non-parametric nearest neighbors.

BibTeX

@inproceedings{azizpour2026video,
  title     = {Video Forensic Self-Descriptions: Leveraging Temporally Distributed Forensic Microstructures for Zero-Shot Detection of {AI}-Generated Videos},
  author    = {Azizpour, Aref and Nguyen, Tai D. and Stamm, Matthew C.},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {39},
  year      = {2026}
}