AI-Generated Video Detection via Perceptual Straightening

🎯 Pitch

AI-generated video trajectories are less straight in neural representation space, and simply measuring this temporal curvature lets a lightweight classifier detect fakes with 97% accuracy—outperforming far heavier models. Remarkably, geometrically straightening synthetic paths by smoothing frames eliminates the detection signal, confirming that generative models fail to replicate the temporal smoothness of natural videos.

1. Executive Summary

This paper introduces ReStraV (Representation Straightening for Video), a method that detects AI-generated videos by analyzing the geometric properties of neural representations. Using a frozen DINOv2 ViT-S/14 vision transformer as a feature space on benchmarks including VidProM, GenVidBench, and Physics-IQ, the approach quantifies temporal curvature (the angle between successive displacement vectors of frame embeddings) and stepwise distance (the magnitude of inter-frame representation change), then trains a lightweight classifier on aggregated statistics of these measures. A two-layer MLP achieves 97.17% accuracy and 98.63% AUROC on VidProM while processing a video end-to-end in approximately 48 ms — substantially outperforming both image-based and video-based state-of-the-art detectors — establishing that the differential straightening of natural versus synthetic video trajectories in self-supervised networks provides a reliable detection signal even for future, unseen generators like Sora and Veo3, though this signal is fundamentally bounded by whether generative models can eventually replicate the temporal smoothness that current systems fail to capture.

2. Context and Motivation

The Core Problem: AI-Generated Videos Are Becoming Indistinguishable from Reality

The fundamental challenge this paper addresses is deceptively simple: as generative AI models produce increasingly realistic synthetic videos, how do we reliably distinguish them from real footage? This isn't merely an academic exercise — it's a problem with escalating real-world stakes. The paper opens by noting that early video generation approaches (GANs, VAEs) struggled with fidelity and temporal coherence, but that "rapidly evolving large-scale foundation models have introduced sophisticated generative techniques" — diffusion models and transformer-based architectures that "can produce near-photorealistic videos from text or initial frames" (Section 1). The gap between real and synthetic video is collapsing, and existing detection methods are not keeping pace.

This matters for several reasons the paper makes explicit:

Content authentication: As synthetic media becomes indistinguishable from real footage, the ability to verify whether a video genuinely depicts events that occurred — or was fabricated — becomes critical for journalism, legal evidence, and public discourse. The paper cites concerns about "malicious manipulation and fabricated visual media" (Section 1).
Disinformation and fraud: The paper references a reported 2137% rise in deepfake-related fraud attempts in the financial sector over three years (Section 8, citing Signicat), making detection not just a theoretical need but an active defense against financial crime.
Epistemic decay: The concluding discussion invokes the concern that without reliable detection, society risks an erosion of shared factual ground — what the authors describe as "the long-term risk of epistemic decay in world models" (Section 8).

The paper positions itself squarely in the detection camp, explicitly contrasting with watermarking approaches that "rel[y] on model operators' goodwill and can be circumvented" (Section 1, citing work on watermarking limitations). Detection, unlike watermarking, does not require cooperation from the generative model's creator — it operates purely on the output.

Why Detecting AI Video Is Harder Than Detecting AI Images

The paper makes a crucial distinction that motivates its video-specific approach. Detecting AI-generated images is relatively well-studied, with a substantial literature on spatial and frequency-domain artifacts (CNNSpot, Gram-Net, FreDect, etc. — all cited in Section 2). But video introduces a dimension that single-frame methods fundamentally miss: temporal consistency. As the paper states:

"Detecting AI videos is more challenging than AI generated image due to temporal consistency requirements that necessitate thorough analysis across frames" (Section 1).

A generated image only needs to look plausible in isolation. A generated video must maintain coherent motion, consistent object identity, and physically plausible dynamics across time. Current generative models struggle with this — the paper cites evidence that even sophisticated models like Sora and VideoPoet "show coherence issues or generate implausible scenarios" (Appendix A), and that "robust temporal consistency and physical plausibility remain significant challenges" despite rapid progress.

This temporal dimension is both the source of the difficulty and, as the paper argues, the key to the solution. Where prior image-based detectors look at individual frames for artifacts (unnatural textures, frequency anomalies), the video-specific challenge demands methods that analyze relationships between frames — how the visual content evolves over time. The paper's central insight is that these temporal relationships, when projected into a neural network's representation space, reveal geometric signatures that distinguish real from synthetic video.

Prior Approaches and Where They Fall Short

The paper surveys two broad categories of existing detection methods, each with specific limitations that motivate the ReStraV approach.

Image-Based Detectors Applied to Video

The first category consists of methods designed for still images, which are sometimes naively applied to video by processing individual frames. The paper lists eight such detectors in its benchmarking (Table 2): CNNSpot, FreDect, Fusing, Gram-Net, GIA, LNP, DFD, and UnivFD. These methods operate on spatial or frequency-domain artifacts within single frames — for example, Gram-Net looks for anomalies in global texture statistics, while FreDect uses frequency masking to highlight generative artifacts, and LNP analyzes gradient patterns characteristic of GAN-generated images.

The limitation is straightforward: these methods are "frame-centric" (Section 2), meaning they treat each frame independently. A deepfake detector tuned to facial artifacts in a single frame has no mechanism to detect that an object flickers in and out of existence across frames, or that motion patterns violate physical laws, or that textures evolve unnaturally over time. The paper's benchmarking confirms this inadequacy empirically: on the VidProM dataset, image-based detectors achieve accuracies ranging from 45% to 64% (Table 2), barely above chance for some methods. Even the best performer (Gram-Net at 63.64% average accuracy) falls far short of what a video-aware method can achieve. Interestingly, these methods show some detectability on specific generators — Gram-Net gets 84.19% on Pika but only 50.46% on ModelScope — indicating that their frame-level artifacts are generator-specific and don't provide a universal signal.

The paper is careful to note that comparing ReStraV to image-based detectors on a video task is "not an even comparison" (Section 7A), but the stark performance gap — ReStraV's 97.06% vs. the best baseline's 63.64% — illustrates that temporal information is not just helpful but essential for robust AI video detection.

Dedicated Video Detectors

The second category encompasses methods specifically designed for video understanding, typically adapted from action recognition architectures. The paper evaluates several: TSM (Temporal Shift Module), I3D, SlowFast, X3D, MVIT-V2, VideoSwin, TPN, UniFormer-V2, and TimeSformer, as well as specialized deepfake video detectors like TALL, NPR, STIL, and DeMamba.

These methods do model temporal information — VideoSwin uses shifted window attention across time, SlowFast uses separate pathways for spatial and temporal processing, TimeSformer applies self-attention to space-time patches. However, the paper identifies several critical shortcomings:

1. They require extensive training on labeled video data. These architectures are typically pre-trained on large action recognition datasets (Kinetics, Something-Something) and then fine-tuned for detection. This creates a dependency on the availability and diversity of training data, which is in constant tension with the rapid evolution of generative models — by the time a detector is trained, new generators have emerged that may produce artifacts the detector has never seen.

2. They struggle with generalization across generative models. The paper's experimental results bear this out vividly. When VideoSwin is trained on four generators and tested on unseen ones (Table 3), accuracy drops from 77.91% (seen generators) to 62.44% (unseen generators) and to 60.70% (future generators like Sora). This is not a small degradation — it's a near-collapse to performance levels that make the detector unreliable in practice. The pattern is not unique to VideoSwin; across the GenVidBench benchmark (Table 4), existing video detectors show extreme inconsistency. I3D achieves 60.11% on CogVideo but 8.15% on MuseV. SlowFast gets 81.63% on some generators but 12.25% on others. These fluctuations suggest that these models are latching onto generator-specific artifacts rather than learning a universal signature of synthetic video.

3. They may overlook a more fundamental signal. The paper argues that existing video detectors, while architecturally capable of temporal reasoning, are essentially learning to detect surface-level anomalies — specific motion artifacts, flicker patterns, or texture inconsistencies characteristic of particular training distributions. The authors hypothesize that "the geometric properties of these trajectories — reflecting the inherent smoothness and predictability of natural dynamics that generative models fail to replicate — offer a more robust signal for detection" (Section 2). In other words, rather than learning what specific artifacts look like, a detector should measure whether the temporal dynamics respect the statistical structure of natural video.

This is a subtle but crucial shift in perspective. Prior detectors ask: "Does this video contain known generative artifacts?" ReStraV asks: "Does this video, when projected into a neural representation space, exhibit the temporal smoothness characteristic of natural visual experience?" The former is brittle to changes in generative technique; the latter should be robust as long as generative models fail to replicate the deep temporal structure of real-world dynamics.

Perceptual Straightening: The Missing Theoretical Framework

The paper's most distinctive intellectual move is connecting the AI video detection problem to the perceptual straightening hypothesis from neuroscience (Hénaff et al., 2019; 2021). This is not a pre-existing detection approach — it's a body of neuroscientific work that the paper repurposes for a new application.

The foundational observation is this: natural videos, when represented as sequences of pixel values, trace highly curved trajectories through pixel space. Even a simple video of an object moving across a static background is a curved path, because the only "straight" video would be a simple cross-fade between two frames. However, when these same videos are processed by the biological visual system (or by certain artificial neural networks), the internal neural representations follow much straighter trajectories. This "perceptual straightening" is hypothesized to reflect predictive coding: the visual system learns the statistical regularities of natural dynamics and transforms inputs into a format where temporal prediction becomes easier — straighter paths are more predictable.

Hénaff et al. demonstrated this in both the macaque visual cortex and in convolutional neural networks, showing that later layers of visual processing produce progressively straighter trajectories for natural videos. The phenomenon has been linked to efficient temporal prediction and has been observed across multiple systems — CNNs, LLMs (Hosseini & Fedorenko, 2023, cited in the paper), and the biological visual pathway.

What ReStraV does is invert the question. Instead of asking whether neural representations straighten natural videos (they do), it asks: do AI-generated videos get straightened to the same degree? The paper's hypothesis is that they do not — and that this differential straightening provides a detection signal.

The motivation for this hypothesis is grounded in the nature of self-supervised learning. DINOv2, the chosen feature extractor, is trained on massive collections of natural images without labels, learning representations that capture the statistical structure of real-world visual data. Videos, even though DINOv2 was not trained on video, inherit these representations frame by frame. Since DINOv2's training data consists exclusively of natural images, its internal geometry is shaped by natural visual statistics. When a naturally recorded video is passed through DINOv2, the frame representations form trajectories that respect this geometry — they tend toward straightness because the model has internalized the predictive structure of real-world dynamics. When an AI-generated video is passed through, the frame representations are novel relative to the model's training distribution. They violate the learned statistical regularities, resulting in trajectories that are more erratic, more curved, less predictable — and therefore measurably different.

This is the paper's theoretical innovation: it frames AI video detection not as artifact detection but as out-of-distribution detection in representation space, where the in-distribution property is temporal smoothness as measured by curvature.

The Specific Gap ReStraV Fills

With this context, the paper's positioning becomes clear. The gap it addresses is threefold:

1. No existing method exploits representation-space trajectory geometry for detection. While related work in video quality assessment has used trajectory analysis (e.g., ST-GREED, cited in Section 2), that work aims to predict perceptual quality for compression or transmission assessment, not to distinguish real from synthetic content. The paper explicitly distinguishes itself: "Unlike related work in video quality assessment that also uses trajectories, our focus is distinctly on detecting synthetic content, regardless of its perceptual quality" (Section 2). An AI-generated video could look visually perfect — high perceptual quality — and still exhibit geometric trajectory anomalies that ReStraV would detect.

2. Existing video detectors achieve moderate performance at high computational cost and poor generalization. The SoTA video detectors require training large architectures on labeled data, struggle with generalization to unseen generators (as documented in Tables 3, 4, and 5), and process video at speeds that make real-time deployment challenging. ReStraV, by contrast, uses a frozen pre-trained encoder with no fine-tuning, trains only a lightweight classifier on pre-computed geometric features, and processes video in approximately 48 ms end-to-end. The efficiency-to-performance ratio is a key part of the contribution.

3. The detection problem lacks a principled theoretical framework. Prior work largely treats AI video detection as an empirical problem: collect data from known generators, train a classifier, hope it generalizes. ReStraV grounds detection in a specific hypothesis — that real videos follow straighter representation-space trajectories because self-supervised networks trained on natural data internalize the predictive structure of real-world dynamics, and that current generative models fail to replicate this structure. This provides not just a method but a falsifiable theoretical prediction: if future generative models learn to produce videos whose neural representations exhibit naturalistic temporal smoothness, ReStraV's signal should weaken. The authors acknowledge this explicitly: "as video models become more and more capable of generating realistic, natural-looking videos, it is possible that future video models may not show the same statistical discrepancies" (Section 8, Limitations). This is not a weakness of the method but an honest characterization of the detection-progress arms race.

How Existing Work Falls Short — A Synthesis

The paper's critique of prior work can be understood along three axes:

Temporal blindness: Image-based detectors (Table 2) achieve near-chance performance on some generators because they ignore the very dimension — time — where generative models currently fail most conspicuously.
Training dependence: Video-based detectors (Tables 3, 4, 5) require expensive training, are fragile to distribution shift, and appear to learn generator-specific artifacts rather than universal signatures of synthetic video.
Theoretical shallowness: Neither category is grounded in a theory of why generative models produce detectable anomalies. They are empirical classifiers that work well on their training distributions and poorly elsewhere.

ReStraV positions itself as addressing all three: it exploits temporal structure (via inter-frame geometric analysis), requires no task-specific training of the feature encoder (DINOv2 is frozen), and is grounded in a neuroscientifically motivated hypothesis about representation-space dynamics. The method is not claimed to be immune to the arms race — the paper is upfront about Goodhart's law and the possibility that future generators might close the curvature gap — but it provides a principled detection framework that performs robustly today and defines the terms under which it might fail tomorrow.

3. Technical Approach

3.1 Reader Orientation

This is primarily a detection method paper whose core idea is that AI-generated videos can be identified by measuring how "straight" their frame-to-frame trajectories are in the representation space of a frozen vision transformer — a property inspired by the neuroscientific finding that biological visual systems straighten natural video trajectories to make temporal prediction easier. The paper builds a complete detection pipeline that extracts frame embeddings using DINOv2, computes per-video geometry statistics (curvature angles and stepwise distances between consecutive frames in representation space), and feeds aggregated statistics of these measures into a lightweight classifier — the whole system requiring no fine-tuning of the feature encoder and processing a video in approximately 48 milliseconds end-to-end.

3.2 Big-Picture Architecture (Diagram in Words)

The ReStraV system has four major components connected in a linear pipeline:

Frame Sampling and Preprocessing — extracts a fixed number of frames from each video at uniform temporal intervals, resizing and normalizing them for the vision transformer. This is the only video-specific preprocessing step; after this, all subsequent processing treats frames as individual images.
Frozen Feature Extractor (DINOv2 ViT-S/14) — a self-supervised vision transformer pre-trained on natural images, used without any fine-tuning. It maps each frame to a high-dimensional embedding vector by concatenating the CLS token and all patch tokens from its final transformer block. This component defines the representation space in which trajectory geometry is measured.
Trajectory Geometry Computation — takes the sequence of frame embeddings and computes two per-step measures: curvature $\theta_i$ (the angle between successive displacement vectors, quantifying how sharply the trajectory bends at each point) and stepwise distance $d_i$ (the Euclidean magnitude of inter-frame change). These are then aggregated into per-video statistics (mean, variance, minimum, maximum) for both curvature and distance, yielding a compact feature vector.
Lightweight Classifier — takes the aggregated geometric features and a small set of individual time-step signals, and produces a binary prediction (natural vs. AI-generated). Multiple off-the-shelf classifiers are evaluated (logistic regression, SVM, random forest, gradient boosting, Gaussian naive Bayes, and a two-layer MLP), with the MLP achieving the best performance. This classifier is the only component trained for the detection task; the feature extractor remains frozen.

Information flows sequentially: raw video → frame sampling → per-frame DINOv2 encoding → per-step curvature and distance computation → per-video statistical aggregation → classifier prediction. There are no feedback loops, no iterative refinement, and no online adaptation — it is a single-pass, deterministic (except for stochastic classifiers like random forest) pipeline.

3.3 Roadmap for the Deep Dive

First, frame sampling and preprocessing (Section 3.4.1): how many frames, over what duration, at what resolution, and why these choices matter for capturing temporal dynamics while maintaining computational efficiency.
Second, the DINOv2 feature extractor (Section 3.4.2): which model variant, which layer, how frame embeddings are constructed from token outputs, and the empirical justification (across 14 visual encoders) for choosing DINOv2 specifically.
Third, the curvature and distance computation (Section 3.4.3): the formal definitions of stepwise distance $d_i$ and curvature $\theta_i$ , how they form a trajectory-level description, and the geometric intuition linking these measures to the perceptual straightening hypothesis.
Fourth, the per-video statistical aggregation (Section 3.4.4): which statistics are computed (mean, variance, min, max for both $d_i$ and $\theta_i^\circ$ ), how they map to an 8-dimensional feature vector, and the evidence from distributional analysis (Figure 5) showing that natural and AI-generated videos separate along these dimensions.
Fifth, the classifier construction (Section 3.4.5): how individual time-step signals are combined with aggregated statistics to form the final 21-dimensional feature vector, which classifier architectures are evaluated, the decision threshold optimization protocol, and the inference-time cost breakdown.
Sixth, the DINOv2 selection analysis (Section 3.4.6): why 14 different visual encoders were compared, the counterintuitive finding that the best perceptual straighteners are not the best detectors, and the interpretation in terms of differential straightening.

3.4 Detailed, Sentence-Based Technical Breakdown

3.4.1 Frame Sampling and Preprocessing

The first stage of the pipeline converts a raw video — potentially of varying length, frame rate, and resolution — into a standardized sequence of image tensors suitable for the DINOv2 vision transformer. Since DINOv2 is an image encoder (not a video model), every frame is processed independently; the temporal structure is captured entirely by the geometric analysis applied to the resulting embedding sequence, not by any architectural mechanism in the feature extractor.

Temporal window and frame count. The paper fixes the analysis to a 2-second window from which $T = 24$ frames are uniformly sampled. This means frames are extracted at a temporal sampling interval of $\Delta t = 2\text{s} / (24 - 1) \approx 87\text{ ms}$ , or equivalently at approximately 11.5 frames per second regardless of the source video's native frame rate. The choice is justified through an ablation study (Appendix B.1, Figure 15) showing that:

A 1-second window (T ≈ 10 frames at 30 FPS with every-3rd-frame sampling) yields AUROC around 92%, which is usable but suboptimal.
A 2-second window (T = 24 frames processed by the standard pipeline) yields AUROC exceeding 96%, representing a strong accuracy-efficiency trade-off.
Longer windows (3–5 seconds, T = 30–50 frames) continue to improve AUROC toward 98%, but at linearly increasing computational cost — more DINOv2 forward passes.

The 2-second, 24-frame configuration is selected as the operating point that "provide[s] an optimal trade-off between high detection accuracy and computational efficiency" (Section 4, main text). Within this 2-second window, the paper further validates that sampling every 3rd frame (at 30 FPS native) outperforms both sparser sampling (every 5th frame) and denser sampling (every frame, or every 2nd frame), with the optimal performance occurring at T = 20–24 frames (Appendix B.1, Figure 15b-c). The degradation at very high sampling density (T = 60, every frame) suggests that excessively fine-grained temporal resolution introduces noise or redundancy that does not improve the geometric signal.

Temporal window position. A separate ablation (Appendix B.2, Figure 16) confirms that the detection performance is robust to where in the video the 2-second window is placed. Sliding the window across a 5-second video with a step of 10 frames (approximately 0.33 seconds) produces accuracy and AUROC values that remain within a narrow band (roughly 95–97% AUROC), with a slight U-shaped trend — marginally higher at the beginning and end of the analyzed range, marginally lower in the middle. This robustness is important for practical deployment: the system does not require temporal alignment or shot boundary detection to work reliably.

Spatial preprocessing. Each sampled frame $x_i$ (for $i \in \{1, \ldots, 24\}$ ) is:

Resized to $224 \times 224$ pixels, matching the input resolution expected by DINOv2 ViT-S/14.
Normalized to the range $[0, 1]$ (Section 4, main text).

These are standard preprocessing steps for vision transformers; no data augmentation, cropping, or temporal alignment is applied.

Design choice: uniform temporal sampling rather than shot-aware sampling. The paper does not attempt to detect or avoid scene cuts. A potential concern is that the perceptual straightening hypothesis is not expected to hold across shot boundaries — a hard cut between two unrelated frames would produce a sharp spike in curvature that could be misinterpreted as a generative artifact. The paper addresses this through two mechanisms. First, the 2-second window inherently reduces the probability of encountering a scene cut. Second, an explicit analysis of scene cut frequency (Appendix C, Figure 17) using the PySceneDetect library on 13,000 AI-generated and 13,000 natural videos shows that the average number of hard cuts in the first 2 seconds is low and comparable across sources — most AI-generated videos from models like Pika, Show-1, ModelScope, and Gen-2 have zero cuts in the first 2 seconds, while natural videos average 0.08 cuts and AI videos range from 0.00 to 0.10. The paper concludes that scene cut frequency is "not a significant confounding factor" (Appendix C).

Why this sampling strategy works for the benchmarks. The paper evaluates on datasets where videos are approximately 2–5 seconds long at 12–30 FPS (Section 4). The 2-second window with 24 uniformly sampled frames captures a substantial fraction of each video's content. For shorter videos near 2 seconds, the entire video is analyzed. For longer videos, any 2-second segment provides a representative sample of the temporal dynamics. The paper hypothesizes that longer videos would further improve performance (Section 4), but the current configuration is sufficient to establish the method's effectiveness.

3.4.2 DINOv2 Feature Extraction

The core representational engine of ReStraV is DINOv2 ViT-S/14 (Oquab et al., 2024), a self-supervised vision transformer pre-trained on a large and diverse collection of natural images. Critically, DINOv2 is used frozen — no fine-tuning, no adaptation, no training on video data whatsoever. The method relies entirely on the representational geometry that DINOv2 learned during its image-level self-supervised pre-training, then repurposes that geometry for temporal analysis.

Model variant and layer selection. The paper uses the ViT-S/14 variant (the "Small" architecture with patch size 14). This choice balances representational quality with computational efficiency — a larger variant (ViT-B, ViT-L) would produce richer embeddings but at higher inference cost. The embeddings are extracted from the final transformer block (referred to as block.11 in the DINOv2 architecture, Section 4). This is the output layer of the vision transformer, representing the highest-level features before any task-specific head. The rationale for using the final layer (rather than intermediate layers) is consistent with the perceptual straightening literature, which shows that later layers of visual processing produce progressively straighter trajectories for natural videos — the straightening effect is most pronounced at the highest representational level.

Embedding construction: CLS token + all patch tokens. For each frame $x_i$ (resized to $224 \times 224$ ), DINOv2 produces:

One CLS (Classify) token of dimension 384 — a global summary embedding that aggregates information across the entire image via the self-attention mechanism.
196 patch tokens of dimension 384 each — corresponding to the $16 \times 16$ grid of non-overlapping $14 \times 14$ pixel patches that tile the input image ( $224 / 14 = 16$ patches per spatial dimension).

The paper concatenates all of these tokens — CLS plus all patch tokens — and flattens them into a single feature vector $z_i \in \mathbb{R}^{75648}$ . The dimensionality is computed as: $(1 \text{ CLS} + 196 \text{ patches}) \times 384 \text{ dim/token} = 197 \times 384 = 75648$ . This is an unusually high-dimensional embedding — substantially larger than using just the CLS token alone — and the paper does not reduce dimensionality via pooling or projection.

Why use all tokens rather than just the CLS token? The paper does not explicitly ablate this choice, but the rationale is implicit in the method's design. The curvature and distance measures operate on the full embedding space, and using all patch tokens means the geometric analysis is sensitive to local temporal inconsistencies — a small region of the frame evolving unnaturally would affect the patch tokens for that spatial location, contributing to the overall trajectory curvature. Using only the CLS token would collapse all spatial information into a single global descriptor, potentially masking localized temporal artifacts. The high dimensionality (75,648) means even subtle frame-to-frame changes in small image regions can contribute to measurable differences in stepwise distance and curvature.

Computational cost of feature extraction. The DINOv2 forward pass for a batch of 8 frames (not the full 24 at once — frames are batched for GPU efficiency) takes 43.6 milliseconds on a single NVIDIA RTX-2080 (Appendix D, Section 6). For a full 24-frame video, this implies approximately $3 \times 43.6 \approx 130.8$ ms of DINOv2 computation split across three batches, though the paper reports the total end-to-end latency as approximately 48 ms (Section 6), suggesting that in practice, frames are processed in parallel or with further optimization. The exact batching strategy is not detailed beyond the "8-frame batch" specification.

Theoretical justification: why DINOv2 specifically? Section 4 of the paper presents a systematic comparison of 14 visual encoders across diverse families — supervised CNNs (AlexNet, VGG-16, ResNet-50, SIN-ResNet-50), self-supervised models (SimCLR-R50, BYOL-R50, CLIP, DINOv2), human visual system-inspired models (Gabor filter bank, LGN-V1), spatio-temporal models (S3d, R(2+1)D, MViT), and an adversarially trained ResNet-50. The key metric is the mean curvature gap $\Delta\theta$ : the difference in average curvature between AI-generated and natural videos when processed by each encoder. A positive gap means AI videos are more curved than natural videos — the desired signal for detection.

DINOv2 produces the largest positive gap: $\Delta\theta = 45.46^\circ$ . This is substantially larger than other self-supervised models (CLIP, BYOL, SimCLR — exact values not tabulated but visible in Figure 2B) and dramatically larger than human visual system-inspired models, which show negative gaps (natural videos are more curved than AI videos in those representation spaces). The finding that HVS-inspired models produce negative deltas is counterintuitive: these models achieve the strongest absolute straightening — they make all trajectories very straight — but they do so indiscriminately for both real and AI-generated videos. The straightening is so aggressive that it erases the differential signal. In contrast, DINOv2 "reduce[s] the curvature of natural videos, which align with their learned priors of real-world statistics, but do[es] not regularize the trajectories of AI-generated videos, which violate these priors" (Section 4). This differential straightening — straightening natural videos substantially more than synthetic ones — is the foundation of the detection signal.

The paper explicitly tests whether absolute straightening capability correlates with detection performance and finds that it does not: the correlation between absolute straightening and the curvature gap is $\rho = -0.13$ with $p = 0.64$ (Section 4). Detection quality is driven not by how well the model straightens trajectories in general, but by how selectively it straightens natural versus synthetic trajectories. This finding is surprising and important — it means that choosing a feature extractor for detection is not about finding the model that produces the straightest trajectories, but about finding the model whose training distribution and inductive biases create the largest representational discrepancy between natural and synthetic inputs.

3.4.3 Curvature and Stepwise Distance Computation

Once the 24 frame embeddings $Z = (z_1, z_2, \ldots, z_{24})$ are extracted, where each $z_i \in \mathbb{R}^{75648}$ , the paper computes two per-step geometric measures that together characterize the shape of the trajectory through representation space. These measures are defined in Equation (1) and are the core signal from which all downstream features are derived.

Stepwise distance $d_i$ : For each consecutive pair of frames $(z_i, z_{i+1})$ , the stepwise distance is the Euclidean ( $L_2$ ) norm of their difference:

$d_i = \|z_{i+1} - z_i\|_2$

for $i = 1, \ldots, T-1$ , where $T = 24$ . This yields a sequence of 23 distance values $\{d_1, d_2, \ldots, d_{23}\}$ .

What it measures: the magnitude of change in the neural representation from one frame to the next. A large $d_i$ indicates that the frames are represented very differently by DINOv2 — perhaps due to large motion, appearance change, or a semantically meaningful event. A small $d_i$ indicates representational similarity — the frames are close in the embedding space. For natural videos, the distribution of $d_i$ values reflects the statistics of real-world temporal change (mostly small changes with occasional large jumps at event boundaries). For AI-generated videos, the distribution may differ because generative models produce frame-to-frame transitions that violate natural temporal statistics.

Curvature $\theta_i$ : For each consecutive triplet of frames $(z_i, z_{i+1}, z_{i+2})$ , the curvature is the angle between two successive displacement vectors. Let $\Delta z_i = z_{i+1} - z_i$ be the displacement from frame $i$ to $i+1$ , and $\Delta z_{i+1} = z_{i+2} - z_{i+1}$ be the displacement from frame $i+1$ to $i+2$ . Then:

$\theta_i = \arccos\left(\frac{\Delta z_i \cdot \Delta z_{i+1}}{\|\Delta z_i\|_2 \cdot \|\Delta z_{i+1}\|_2}\right)$

for $i = 1, \ldots, T-2$ . This yields a sequence of 22 curvature values $\{\theta_1, \theta_2, \ldots, \theta_{22}\}$ . The curvature is typically converted from radians to degrees:

$\theta_i^\circ = \theta_i \times \frac{180}{\pi}$

What it computes: the angle $\theta_i$ is the angle at vertex $z_{i+1}$ formed by the three consecutive embeddings $z_i$ , $z_{i+1}$ , and $z_{i+2}$ . If $\theta_i = 0^\circ$ , the three points are collinear and the trajectory does not bend at all — it is perfectly straight at that step. If $\theta_i = 180^\circ$ , the trajectory makes a complete reversal — the direction of change from $i$ to $i+1$ is exactly opposite to the direction from $i+1$ to $i+2$ . Intermediate values represent varying degrees of bending.

Why the arccos form? The cosine similarity between $\Delta z_i$ and $\Delta z_{i+1}$ captures whether successive changes in representation space point in similar directions. If two displacement vectors point in the same direction (cosine similarity = 1), the trajectory continues straight (arccos(1) = 0). If they point in opposite directions (cosine similarity = -1), the trajectory makes a sharp reversal (arccos(-1) = $\pi$ radians = 180°). The arccos converts the bounded [-1, 1] similarity into an unbounded [0°, 180°] angle, which is more interpretable: a trajectory with mean curvature near 0° is essentially straight, while a trajectory with mean curvature near 90° is highly erratic.

Why not use a simpler measure of trajectory shape? Alternatives could include:

Turning rate (change in direction per unit time) would capture similar information but requires numerical differentiation which is sensitive to noise.
Fractal dimension or path length would measure overall complexity but would not distinguish between a smooth curved path (low local curvature, high global curvature) and an erratic jagged path (high local curvature).
Direct cosine similarity between non-consecutive frames would capture longer-range predictability but at the cost of temporal resolution.

The curvature measure is attractive because it is local (each $\theta_i$ depends only on three consecutive frames), rotation-invariant (depends only on the angle, not the absolute orientation of the displacement vectors), and directly interpretable through the perceptual straightening lens — the hypothesis is precisely about the angles between successive representational changes.

The geometric picture. Figure 3 illustrates the computation: the three embeddings $z_i$ , $z_{i+1}$ , $z_{i+2}$ form a dashed trajectory through the high-dimensional space. The orange vectors $\Delta z_i$ and $\Delta z_{i+1}$ represent the displacements, with lengths $d_i$ and $d_{i+1}$ . The green angle $\theta_i$ at point $z_{i+1}$ is the curvature — the angle by which the trajectory turns at that step. A "straight" trajectory would have all $\theta_i \approx 0^\circ$ ; a "curved" trajectory would have some $\theta_i$ substantially larger.

The perceptual straightening hypothesis as applied here. The core claim is that for a frozen DINOv2 encoder:

Natural videos produce trajectories with relatively low mean curvature $\mu_\theta$ — they are "straightened" by the representation space.
AI-generated videos produce trajectories with higher mean curvature — they are not straightened, or are straightened less, because they violate the learned priors.

However, the paper observes a more nuanced pattern (Section 8, Discussion): natural videos have lower mean curvature but higher variance in curvature. The authors connect this to prior work showing that "temporal transitions in natural videos latent representations follow highly sparse distributions" — most of the time, natural videos change very little (producing near-zero curvature), but occasionally make large jumps (producing sharp curvature spikes, possibly at scene cuts or event boundaries). The distribution is therefore characterized by a low mean and a high variance. AI-generated videos, by contrast, show more consistently high curvature — they deviate from natural statistics across many frames, not just at event boundaries.

Output. The curvature and distance computation yields two time series per video:

$\{d_i\}_{i=1}^{23}$ : 23 stepwise distance values.
$\{\theta_i^\circ\}_{i=1}^{22}$ : 22 curvature values (in degrees).

These raw time series are the input to the statistical aggregation step.

3.4.4 Statistical Aggregation into Per-Video Feature Vectors

The raw curvature and distance time series are high-dimensional and video-specific (each video produces 45 values: 23 distances + 22 curvatures). Training a classifier directly on these raw sequences would require temporal alignment and would be sensitive to video length. Instead, the paper computes aggregated statistics that summarize each video's trajectory geometry into a compact, fixed-dimensional feature vector.

The 8-dimensional aggregated feature vector. For each video, the paper computes the first four descriptive moments of both the distance and curvature sequences:

For distance $\{d_i\}$ : mean $\mu_d$ , minimum $\min d$ , maximum $\max d$ , and variance $\sigma_d^2$ .
For curvature $\{\theta_i^\circ\}$ : mean $\mu_\theta$ , minimum $\min \theta$ , maximum $\max \theta$ , and variance $\sigma_\theta^2$ .

Formally:

$\mu_d = \frac{1}{T-1}\sum_{i=1}^{T-1} d_i, \quad \sigma_d^2 = \frac{1}{T-1}\sum_{i=1}^{T-1} (d_i - \mu_d)^2$

and analogously for the curvature (using $\theta_i^\circ$ and summing over $T-2$ values).

The aggregated feature vector is therefore:

$\left[\mu_d, \min d, \max d, \sigma_d^2, \mu_\theta, \min \theta, \max \theta, \sigma_\theta^2\right] \in \mathbb{R}^8$

What these statistics capture:

$\mu_d$ and $\mu_\theta$ : the central tendency — on average, how much do frames change, and how curved is the trajectory? Higher mean curvature indicates less straightening.
$\min d$ and $\max d$ , $\min \theta$ and $\max \theta$ : the range — what are the most extreme values? These capture outlier frames (e.g., a scene cut producing a very sharp curvature spike).
$\sigma_d^2$ and $\sigma_\theta^2$ : the variability — how consistent is the frame-to-frame behavior? The paper finds that natural videos show higher variance in curvature despite lower mean curvature (Section 8, Discussion), suggesting that the variance captures the sparse-event structure of natural temporal dynamics (mostly straight, occasionally sharp turns).

Why these four moments? The first four statistical moments (mean, variance, skewness, kurtosis) form a standard summary of any distribution. The paper uses mean, variance, min, and max — a slightly different set than the classical four moments (min/max replace skewness/kurtosis). Min and max are chosen presumably because they are more interpretable (capturing extreme values directly) and because they are more robust to outliers than higher-order moments. The paper does not ablate this specific choice of statistics, but the permutation feature importance analysis (Appendix A.3, Figure 11) confirms that the aggregated statistics — particularly $\mu_\theta$ — are by far the most important features for classification.

Empirical validation of the aggregated features. Section 5 presents a large-scale distributional analysis using 50,000 AI-generated videos (10,000 each from Pika, VideoCrafter2, Text2Video-Zero, ModelScope, and Sora) and 50,000 natural videos from DVSC2023. Figure 5 shows the distributions of all 8 aggregated features, and the separation is visually striking:

For distance-based features (top row), the distributions of $\mu_d$ , $\min d$ , $\max d$ , and $\sigma_d^2$ all show clear shifts between natural and AI-generated videos. The AI-generated distributions tend toward higher values, indicating more inter-frame change.
For curvature-based features (bottom row), the separation is even more pronounced. The $\mu_\theta$ distribution for AI-generated videos is shifted substantially rightward (higher mean curvature), while the $\sigma_\theta^2$ distribution shows natural videos with higher variance — exactly the "low mean, high variance" pattern described in Section 8.

Statistical tests confirm these differences: a two-sample t-test comparing mean per-video $\mu_\theta$ between natural and AI-generated videos yields $t = 44.02$ with $p \approx 0$ (effectively zero), and an ANOVA comparing feature distributions across different AI generators and natural videos yields $F = 18598.17$ with $p \approx 0$ (Section 5). These are extremely strong statistical separations, confirming that the aggregated features provide a reliable discriminative signal even before any classifier is trained.

3.4.5 Classifier Construction and Training

With the geometric features computed and aggregated, the final component of ReStraV is a supervised classifier that maps the feature vector to a binary prediction (natural vs. AI-generated). The paper evaluates multiple off-the-shelf classifiers and selects the best performer based on held-out test performance.

The full 21-dimensional feature vector. For the classifier, the paper augments the 8 aggregated statistics with a subset of the individual time-step signals (Section 6). Specifically, the classifier's input $y \in \mathbb{R}^{21}$ concatenates:

7 distance values: $d_1, d_2, \ldots, d_7$ — the stepwise distances for the first seven frame transitions.
6 curvature values: $\theta_1^\circ, \theta_2^\circ, \ldots, \theta_6^\circ$ — the curvatures for the first six frame triplets.
8 aggregated statistics: $\mu_d, \min d, \max d, \sigma_d^2, \mu_\theta, \min \theta, \max \theta, \sigma_\theta^2$ .

Why include individual time-step signals? The aggregated statistics capture global properties of the trajectory, but the earliest frames may contain particularly informative signals — for instance, many generative models produce an initial frame that is relatively coherent, then exhibit increasing temporal inconsistency as the video progresses. The first few $d_i$ and $\theta_i$ values capture this early-trajectory behavior. The permutation feature importance analysis (Appendix A.3, Figure 11) confirms that aggregated statistics dominate importance (with $\mu_\theta$ being by far the most important feature, followed by $\sigma_\theta^2$ and other aggregated statistics), while individual time-step signals provide "complementary but less critical information." The inclusion of these individual signals adds dimensionality without substantially increasing computational cost, since they are already computed during trajectory analysis.

Training data and split. The classifier training uses the same dataset described in Section 5: 50,000 AI-generated videos (balanced across five generators: Pika, VideoCrafter2, Text2Video-Zero, ModelScope, and Sora) and 50,000 natural videos from DVSC2023. The split is stratified 50/50 train/test, with class priors identical in both subsets (50% natural, 50% AI-generated) and each subset balanced among the five AI models (Section 6). This means the training set contains 25,000 natural and 25,000 AI-generated videos (5,000 from each generator), and the test set mirrors this composition.

Classifiers evaluated. The paper evaluates six off-the-shelf models (Section 6, Table 1), with no feature engineering or extensive hyperparameter search beyond "a 3-fold grid/random sweep":

Logistic Regression (LR): a linear classifier that learns a hyperplane separating the two classes in the 21-dimensional feature space. Serves as a baseline for whether the features are linearly separable.
Gaussian Naive Bayes (GNB): assumes features are conditionally independent given the class and follow Gaussian distributions. Fast but likely suboptimal given feature correlations.
Random Forest (RF): ensemble of 400 decision trees with maximum depth ≤ 6. Captures non-linear feature interactions and provides robustness through bagging.
Gradient Boosting (GB): sequential ensemble of 200 decision trees with learning rate 0.1. Builds classifiers iteratively, with each tree correcting the errors of the previous ensemble.
RBF-kernel SVM: support vector machine with radial basis function kernel, calibrated by Platt scaling to produce probability estimates. Captures complex non-linear decision boundaries.
Two-layer MLP: a small feedforward neural network with architecture 64 → 32 (64 hidden units in the first layer, 32 in the second, presumably with ReLU activations though the paper does not specify the activation function). This is the "headline" ReStraV classifier.

Decision threshold optimization. For each classifier, the decision threshold $\tau^*$ is optimized on the training set to maximize the F1-score (the harmonic mean of precision and recall). The optimized threshold is then applied unchanged to the test set. This is a standard protocol that prevents inflated test performance from threshold tuning on the test data itself. The paper does not report the specific threshold values.

Classifier performance (Table 1). The results on the 50k/50k balanced test set:

Model	Accuracy	Balanced Accuracy	Specificity	Precision (Gen)	Recall (Gen)	F1 (Gen)	AUROC	Time (ms)
SVM	85.23	85.78	86.42	96.93	85.04	90.62	93.27	1183.94
GNB	86.64	84.43	81.12	95.94	87.72	91.68	92.05	44.53
LR	89.02	88.86	88.53	97.54	89.17	93.12	95.26	43.97
GB	92.83	92.31	91.57	98.25	93.16	95.63	97.85	48.59
RF	94.24	88.67	80.37	96.13	97.05	96.53	98.03	48.14
MLP	94.17	94.19	94.11	98.88	94.14	96.48	98.63	48.12

The MLP achieves the highest accuracy (94.17%), balanced accuracy (94.19%), precision (98.88%), and AUROC (98.63%), making it the recommended ReStraV classifier. Random forest achieves slightly higher recall (97.05%) and the second-highest AUROC (98.03%), but at the cost of lower specificity (80.37%) — meaning it has a higher false positive rate, misclassifying 19.63% of natural videos as AI-generated. Gradient boosting achieves the second-highest accuracy overall (92.83%) and strong precision (98.25%). The linear models (LR, SVM) underperform, confirming that the feature space is not linearly separable and that non-linear interactions between features matter for optimal classification.

Why the MLP? The paper does not provide an extensive justification, but the pattern in the results suggests that the MLP's combination of high accuracy and balanced specificity/recall makes it the most reliable classifier. The 64 → 32 architecture is small enough to avoid overfitting (only a few thousand parameters) but has sufficient capacity to capture the non-linear decision boundaries visible in Figure 10 (Appendix A.2). The MLP's inference time (48.12 ms total, including DINOv2) is nearly identical to GNB (44.53 ms) — the DINOv2 forward pass (43.6 ms) dominates the computational cost, making the choice of classifier essentially free in terms of latency.

Inference-time cost breakdown. The paper reports end-to-end latency as $\text{latency} = T_{\text{DINOv2}} + T_{\text{clf}}$ , where $T_{\text{DINOv2}} = 43.6$ ms for the DINOv2 forward pass (8-frame batch), and $T_{\text{clf}}$ is the classifier-specific time. For the MLP, $T_{\text{clf}} \approx 48.12 - 43.6 = 4.52$ ms — negligible compared to feature extraction. The total latency of approximately 48 ms per video is fast enough for near-real-time applications and is orders of magnitude faster than training and running a dedicated video detection model like VideoSwin or TimeSformer.

Design choice: frozen encoder + lightweight classifier vs. end-to-end training. The paper deliberately avoids fine-tuning DINOv2 or training a larger end-to-end detection model. This choice is motivated by several considerations:

Computational efficiency: Fine-tuning a vision transformer on video data would require substantial GPU resources and time. The frozen-encoder approach requires only computing forward passes (no gradient computation for DINOv2) and training a tiny classifier on pre-computed features.
Generalization: Fine-tuning risks overfitting to the specific generators in the training set, potentially reducing generalization to unseen generators. The frozen DINOv2 encoder has not been exposed to any AI-generated training data, so its representational geometry is determined entirely by natural image statistics — making the geometric signal more likely to generalize.
Reproducibility: The frozen-encoder approach is easily replicated. Anyone can download DINOv2, extract features, and train a classifier. No specialized hardware, large-scale training infrastructure, or access to generative model training data is required.

Feature importance analysis (Appendix A.3, Figure 11). The permutation feature importance analysis for the MLP reveals that $\mu_\theta$ (mean curvature) is "unequivocally the most critical feature," with an importance "more than double that of the next most influential feature" (curvature variance $\sigma_\theta^2$ ). The eight aggregated statistical features occupy the top importance ranks, while individual time-step signals have "a much smaller impact." This strongly validates the core hypothesis: the overall straightness of the trajectory — as captured by mean curvature — is the primary discriminative signal.

ROC and confusion matrix analysis (Appendix A.2, Figure 9). The MLP's ROC curve shows an AUROC of 98.63%, with the curve hugging the top-left corner — indicating near-perfect separation between the two classes across a wide range of decision thresholds. The normalized confusion matrix shows:

Natural videos: 93.3% correctly classified as natural, 6.7% false positive rate (misclassified as AI-generated).
AI-generated videos: 94.5% correctly classified as AI-generated, 5.5% false negative rate (misclassified as natural).

These error rates are symmetric and low, confirming that the classifier does not exhibit strong bias toward either class.

3.4.6 Why DINOv2? — The Encoder Selection Analysis

The choice of DINOv2 as the feature extractor is not arbitrary — it emerges from a systematic comparison of 14 visual encoders spanning diverse architectural families and training paradigms (Section 4, Figure 2B). This analysis is central to the paper's theoretical argument because it demonstrates that detection capability is not simply a function of representational quality or absolute straightening, but of differential straightening — the tendency to straighten natural video trajectories substantially more than AI-generated ones.

Encoders compared:

Supervised CNNs: AlexNet, VGG-16, ResNet-50, SIN-ResNet-50 (a ResNet-50 trained on a texture-debiased version of ImageNet to increase shape bias rather than texture bias).
Self-supervised models: SimCLR-R50, BYOL-R50, CLIP (ViT-based), DINOv2 (ViT-S/14).
Human visual system-inspired models: a Gabor filter bank (mimicking V1 simple cell receptive fields) and the LGN-V1 model from Hénaff et al. (2019) — these are not trained neural networks but hand-designed or biologically constrained models that approximate early visual processing.
Spatio-temporal models: S3d, R(2+1)D, MViT — these are video architectures trained on action recognition datasets (e.g., Kinetics) and are the only models in the comparison that have been explicitly trained on video data.
Adversarially trained: an $\ell_\infty$ -adversarially trained ResNet-50, which is designed to be robust to imperceptible input perturbations and may have different representational geometry as a result.

The metric: mean curvature gap $\Delta\theta$ . For each encoder, the paper computes the average curvature (across frames and across videos) for AI-generated videos and for natural videos. The difference between these averages — $\Delta\theta = \bar{\theta}_{\text{AI}} - \bar{\theta}_{\text{natural}}$ — is the curvature gap. A positive gap means AI videos are more curved on average; a larger positive gap means better separability.

Results (Figure 2B):

Self-supervised models (green bars) show the largest positive gaps, with DINOv2 achieving $\Delta\theta = 45.46^\circ$ — the highest of any encoder. CLIP, BYOL, and SimCLR also show positive gaps, but substantially smaller than DINOv2.
Human visual system-inspired models (red bars) show negative gaps — in these representation spaces, natural videos are actually more curved than AI-generated videos. The LGN-V1 model and Gabor filter bank aggressively straighten all trajectories, but do so more for AI videos than for natural ones, eliminating the differential signal.
Supervised CNNs and spatio-temporal models show small positive gaps (roughly 5–20°), indicating some differential straightening but much weaker than DINOv2.
Adversarially trained ResNet-50 shows a small positive gap, comparable to standard ResNet-50.

The counterintuitive finding: good perceptual straighteners are not good detectors. The paper computes the correlation between absolute straightening capability (how much an encoder reduces curvature relative to pixel space) and the curvature gap $\Delta\theta$ . The correlation is $\rho = -0.13$ with $p = 0.64$ — essentially zero and not statistically significant. This means that an encoder's tendency to produce straight trajectories overall does not predict its ability to separate natural from AI-generated videos. The HVS-inspired models are the strongest absolute straighteners — they produce the straightest trajectories for all inputs — but they are the worst detectors because they erase the differential signal. DINOv2 is a moderate straightener (it straightens natural videos substantially but not as aggressively as the biological models) but a superb detector because it straightens selectively.

Interpretation: differential straightening as distributional mismatch detection. The paper interprets this result through the lens of training data distribution. DINOv2 was trained on a massive corpus of natural images using self-supervised objectives (student-teacher distillation with centering and sharpening). During this training, DINOv2 learned representations that capture the statistical structure of natural visual data — textures, shapes, object configurations, and scene layouts that occur in real photographs. When a natural video is passed through DINOv2 frame by frame, the representations of consecutive frames tend to be consistent with these learned statistics, producing trajectories that are relatively straight because the model "expects" the types of changes that occur between frames of natural video. When an AI-generated video is passed through, the frame-to-frame changes violate these expectations — the model has not learned to predict the specific types of temporal artifacts, texture evolution, or motion inconsistencies that generative models produce. The result is more erratic trajectories with higher curvature.

This interpretation also explains why spatio-temporal models (S3d, R(2+1)D, MViT) do not perform as well: these models were trained on video data (action recognition) and have learned to be invariant to some types of temporal variation, potentially smoothing over the very inconsistencies that provide the detection signal. Their training objective (classifying actions) encourages temporal robustness, which may partially erase the geometric signatures of synthetic video.

Why DINOv2 specifically? Among self-supervised models, DINOv2 outperforms CLIP, BYOL, and SimCLR. The paper does not explore why in depth, but several factors may contribute: DINOv2's training recipe (which includes multi-crop augmentation, momentum teacher, and carefully tuned hyperparameters) produces highly structured representations that are particularly sensitive to distribution shift. DINOv2 has been shown to excel at out-of-distribution detection and dense prediction tasks, suggesting that its representations capture fine-grained statistical regularities that other SSL methods may miss. The ViT architecture (as opposed to ResNet-based SimCLR and BYOL) may also contribute, as the self-attention mechanism can capture long-range dependencies that are relevant for detecting frame-to-frame inconsistencies.

Summary of Key Design Choices

The ReStraV pipeline is characterized by a set of deliberate design choices, each with specific justifications:

Frozen DINOv2, no fine-tuning: avoids overfitting to training generators, preserves the natural-image prior, enables zero-shot generalization to unseen generators, and ensures computational efficiency (no gradient computation for the feature extractor).
All patch tokens concatenated, not just CLS: captures local temporal inconsistencies that would be averaged out in a global embedding; the 75,648-dimensional embedding space amplifies subtle frame-to-frame changes.
Curvature measured as arccos of consecutive displacement cosine similarity: local, rotation-invariant, directly interpretable through the perceptual straightening lens, and sensitive to the type of temporal irregularity that generative models exhibit.
First four statistical moments (mean, variance, min, max) as aggregators: compact (8 dimensions), interpretable, and validated by both distributional analysis (Figure 5) and permutation feature importance (Figure 11) as capturing the primary discriminative signal.
2-second window with 24 frames, every-3rd-frame sampling: empirically validated as the accuracy-efficiency sweet spot through ablation studies (Appendix B).
Lightweight MLP (64 → 32) rather than larger architectures: sufficient capacity for the non-linear decision boundaries in the feature space, negligible inference cost relative to DINOv2, and no risk of overfitting given the small number of parameters relative to the 50k training samples.
No temporal alignment, no shot detection, no video-specific preprocessing beyond frame extraction: the method is deliberately simple and general, relying on the learned representational geometry rather than hand-crafted video analysis techniques.

4. Key Insights and Innovations

Innovation 1: Repurposing Perceptual Straightening as a Detection Signal, Not a Biological Phenomenon

The field of AI-generated content detection has operated largely as an empirical engineering discipline: collect data from known generators, train classifiers to spot artifacts, and hope for generalization. When generalization fails — as it consistently does in Tables 3–5 — the response is to collect more data, train bigger models, or engineer more sophisticated architectures. The underlying conceptual framework remains unchanged: detection is about learning to recognize the outputs of specific generative processes.

ReStraV makes a fundamentally different move. It takes a neuroscientific hypothesis — that biological visual systems straighten natural video trajectories to support temporal prediction (Hénaff et al., 2019, 2021) — and repurposes it as a detection principle. The insight is not "neural networks exhibit perceptual straightening" (that was already known, and the paper confirms it's not even a good predictor of detection capability, given the near-zero correlation ρ = -0.13 between absolute straightening and the curvature gap). The insight is that differential straightening — the selectivity with which a representation straightens natural versus synthetic trajectories — constitutes a measurable signal of synthetic origin. This transforms a descriptive neuroscientific finding into a prescriptive detection framework.

What makes this intellectually distinctive is the shift in analytical stance. Prior detectors ask a classification question: "Does this video belong to the distribution of known generated content?" ReStraV asks a geometry question: "Does this video's trajectory through a fixed representation space respect the temporal statistics that the representation was built to capture?" The detection signal comes not from learning what synthetic videos look like, but from measuring a mismatch between the video's dynamics and the representation's expectations. The representation — DINOv2, frozen and never exposed to AI-generated content — serves as a normative model of natural temporal structure, and any video whose frame-to-frame transitions violate this structure is flagged as suspicious.

This reframing has profound practical consequences that the paper demonstrates but does not fully articulate. Because the detection criterion is "deviation from natural temporal smoothness" rather than "similarity to known generative artifacts," the method should generalize to any generative model that fails to produce naturalistically smooth trajectories — including future, unseen generators. The paper validates this empirically with zero-shot generalization to Sora (Table 3: 80.05% accuracy despite no Sora videos in training) and Veo3 (83.2% accuracy in zero-shot evaluation), but the conceptual point is deeper: the generalization ceiling is determined not by the diversity of training generators, but by whether generative models eventually learn to produce temporally coherent representations. This is a falsifiable boundary condition that defines the terms of the detection-generator arms race in a way that artifact-based methods cannot.

The comparison to prior work is stark. Image-based detectors (CNNSpot, Gram-Net, FreDect, etc.) achieve 45–64% accuracy on VidProM (Table 2) because they ignore temporal structure entirely. Dedicated video detectors (VideoSwin, TimeSformer, I3D) achieve higher performance on seen generators but collapse on unseen ones (VideoSwin drops from 77.91% to 62.44% in the unseen-generator condition, Table 3) because they learn generator-specific temporal artifacts. ReStraV's differential straightening framework occupies a third category: it exploits temporal structure without learning what specific artifacts look like, achieving both high performance and robust generalization from a single representational principle.

This is a fundamental shift, not an incremental refinement. It doesn't improve an existing detector architecture; it replaces the detection paradigm with one grounded in a theory of what makes natural video natural — and measures departures from that theory in representation space.

Innovation 2: The Differential Straightening Phenomenon as an Empirical Discovery Distinct from Absolute Straightening

Prior work on perceptual straightening (Hénaff et al., 2019, 2021; Niu et al., 2024) established that neural networks and biological visual systems produce straighter trajectories for natural videos than pixel space does, and that this straightening increases across processing stages. The implicit assumption — never explicitly stated but visible in how the finding is discussed — is that straightening is a monolithic property: a good visual system straightens trajectories, and better systems straighten them more.

ReStraV's encoder comparison (Section 4, Figure 2B) shatters this assumption by demonstrating that absolute straightening capability and detection-relevant straightening are not merely uncorrelated — they can be inversely related. The HVS-inspired models (Gabor filter bank, LGN-V1) achieve the strongest absolute straightening, making all trajectories — both natural and AI-generated — extremely straight. But this aggressive straightening erases the very signal that ReStraV exploits: the relative difference in curvature between natural and synthetic trajectories. The HVS models show negative curvature gaps (Δθ < 0), meaning natural videos are actually more curved than AI-generated ones in these representation spaces. DINOv2, by contrast, is a moderate absolute straightener but produces the largest positive curvature gap (Δθ = 45.46°), straightening natural videos substantially more than synthetic ones.

This is an empirical discovery with theoretical implications, not merely a feature-engineering choice. It reveals that representational straightening is not a single phenomenon but has at least two distinct components: absolute straightening (the overall reduction in curvature relative to pixel space) and selective straightening (the degree to which straightening is specific to inputs that match the training distribution). These components are dissociable — models can be high in one and low in the other — and only selective straightening provides a detection signal.

The finding recasts the encoder selection problem. Rather than asking "which model produces the straightest trajectories?" (the natural question if one naïvely imports the perceptual straightening framing), the relevant question is "which model's training distribution and inductive biases create the largest representational discrepancy between natural and synthetic inputs?" DINOv2 wins not because it's the best visual model by standard metrics, but because its self-supervised training on natural images, combined with the ViT architecture's sensitivity to fine-grained statistical regularities, produces representations that are tightly tuned to natural statistics — and therefore most disrupted by synthetic inputs that violate those statistics.

This is significant beyond detection because it suggests a general principle: self-supervised models trained exclusively on natural data may serve as generic out-of-distribution detectors for synthetic content, with the detection signal emerging from the geometry of their representational trajectories rather than from any explicit training for detection. The paper doesn't explore this generalization, but the logic extends naturally: any generative model whose outputs violate the temporal statistics learned by a natural-data-trained SSL encoder should produce measurably different trajectory geometries. This positions ReStraV not just as a specific detection method but as an instance of a broader class of approaches that use representation-space geometry for authenticity verification.

This innovation is fundamental in its implications, even though the paper's direct contribution is the empirical demonstration. It redefines what "good features for detection" mean — not features that make real and fake look different after training, but features where real and fake look different by default, because the features were built to model real data and fake data fails to conform.

Innovation 3: Curvature as a Single Diagnostic Statistic That Subsumes Complex Temporal Modeling

The video detection literature has developed increasingly sophisticated architectures for temporal reasoning: 3D convolutions (I3D, S3d), space-time attention (TimeSformer, VideoSwin), multi-pathway designs (SlowFast), and specialized deepfake detectors (DeMamba, TALL, STIL). These models devote thousands or millions of parameters to learning what temporal anomalies look like, and they require large-scale video datasets for training.

ReStraV demonstrates that for the specific task of AI video detection, a single scalar statistic — the mean temporal curvature μ_θ in a frozen DINOv2 representation space — provides a detection signal that rivals or exceeds these complex architectures, particularly in generalization settings. The permutation feature importance analysis (Appendix A.3, Figure 11) shows that μ_θ alone accounts for more than double the importance of the next most influential feature (σ_θ^2, the variance of curvature). The remaining 19 features in the classifier's 21-dimensional input — distance statistics, individual time-step signals — provide complementary but substantially smaller contributions.

This is not just an efficiency claim. It's a diagnostic insight: the geometric property that distinguishes natural from AI-generated video is not a complex pattern requiring deep architectural reasoning to extract. It is a simple, interpretable, physically meaningful measure — how sharply the representation-space trajectory bends, on average — that can be computed from frozen frame embeddings with no learned temporal parameters whatsoever. The "temporal modeling" is done entirely by the curvature computation, which has zero learnable parameters and requires only the sequence of frame embeddings that DINOv2 produces independently per frame.

The contrast with prior approaches illuminates what was actually being learned. When VideoSwin achieves 77.91% accuracy on seen generators (Table 3), it is presumably learning to recognize specific motion artifacts, flicker patterns, or texture evolution signatures characteristic of those generators' outputs — hence its collapse to 62.44% on unseen generators that produce different artifacts. ReStraV's mean curvature, by contrast, is not learning any artifact signature. It is measuring a fundamental geometric property that reflects whether the video's temporal dynamics are consistent with the representation's learned expectations. This property happens to differ between real and synthetic video, but the difference is not generator-specific — it is a structural consequence of the gap between natural video statistics and current generative models' temporal coherence.

The practical implication is profound: temporal modeling for detection may not require temporal training. A frozen image encoder, never exposed to video during training, can serve as the representational substrate for temporal analysis if the analysis focuses on geometric properties of the embedding trajectory rather than learned spatiotemporal features. This decouples the feature extraction problem (solved by large-scale image pretraining) from the temporal analysis problem (solved by simple geometric computation), eliminating the need for video-specific pretraining and its associated data requirements, computational costs, and generalization fragility.

This is a fundamental simplification, not an incremental architecture improvement. It doesn't propose a better temporal module; it shows that a much simpler temporal analysis — one that doesn't involve learning at all — suffices when paired with an appropriately structured representation space. The finding invites a rethinking of what temporal modeling means in detection contexts: perhaps the heavy lifting should be done by the representational geometry learned during image pretraining, with temporal analysis reduced to measuring how well a sequence respects that geometry.

Innovation 4: The Recognition That Detection Performance Depends on Differential, Not Absolute, Straightening — With the Implication That Stronger Straighteners Can Be Worse Detectors

The field's natural instinct, when confronting a new signal like representational curvature, would be to maximize that signal: find the model that produces the straightest trajectories, or the largest absolute difference between real and fake, and build a detector around it. ReStraV's encoder analysis (Figure 2B, Section 4) shows that this instinct would lead to the wrong choice. The best absolute straighteners — the HVS-inspired models — are the worst detectors, producing negative curvature gaps that would yield below-chance classification. The best detector — DINOv2 — is a moderate straightener that produces highly asymmetric straightening: substantial for natural videos, minimal for synthetic ones.

This is a counterintuitive negative result with important methodological implications. It establishes that the detection signal is not the curvature itself but the curvature differential, and that this differential can be destroyed by representations that are "too good" at straightening — that is, representations that straighten everything equally. The Gabor filter bank and LGN-V1 model, by imposing strong, generic smoothness constraints derived from early visual processing, produce trajectories that are uniformly straight regardless of input, erasing the very inconsistency that makes AI-generated video detectable.

This insight reframes the feature extractor selection problem as an optimization over selectivity, not sensitivity. The goal is not to maximize the curvature of AI-generated trajectories (making them as detectably "wrong" as possible) but to find representations where natural trajectories are straightened while synthetic ones are not — a property that depends on the match between the representation's training distribution and the statistics of natural video, not on any absolute measure of representational quality.

The broader significance extends beyond this paper's specific application. It suggests a general principle for using pretrained models as anomaly detectors: the model should be well-matched to the target distribution but not overtrained or over-regularized to the point of generalizing to anomalous inputs. A model that is "too robust" — that processes all inputs through the same smoothness constraints regardless of their origin — will fail as a detector because it will normalize away the anomalies. This principle is visible in the poor performance of adversarially trained models (which are designed to be robust to input perturbations and might therefore smooth over generative artifacts) and spatio-temporal models (which are trained to be invariant to many types of temporal variation). The best detector is one whose representations are brittle to distribution shift in exactly the way that current generative models produce it.

This is a fundamental conceptual contribution rather than an incremental technical one. It changes how one thinks about model selection for detection: from "choose the best model" to "choose the model whose failure modes align with the anomalies you want to detect." The paper provides both the empirical demonstration (Figure 2B) and the analytical framework (differential vs. absolute straightening) to operationalize this principle, establishing a new criterion — selective straightening — that future work can use to evaluate candidate feature extractors for detection tasks.

The evidence for this innovation is concentrated in Figure 2B and the accompanying correlation analysis (ρ = -0.13, p = 0.64 between absolute straightening and the curvature gap), but its implications ripple through the entire paper: it explains why DINOv2 was chosen (not just because it works, but because of a principled property), it explains why spatio-temporal models underperform (they're too temporally invariant), and it defines a research program for improving detection (find representations with larger differential straightening, not representations that straighten more aggressively).

Innovation 5: ReStraV Defines the Terms of the Detection-Generation Arms Race Through a Falsifiable Hypothesis

Most detection papers conclude with a performance claim: "our method achieves state-of-the-art accuracy on benchmark X." These claims are inherently backward-looking — they demonstrate effectiveness against current generators but say nothing about future ones, and they provide no framework for predicting when or how the method might fail.

ReStraV makes a different kind of claim. By grounding detection in a specific mechanism — differential straightening of representation-space trajectories — it implicitly defines the conditions under which the method would stop working: generative models would need to produce videos whose frame-to-frame transitions, when encoded by a natural-image-trained vision transformer, exhibit the same temporal smoothness (low mean curvature, high curvature variance) as natural videos. This is a concrete, testable, falsifiable prediction about what future generators would need to achieve to evade detection.

The paper is unusually candid about this, acknowledging in the Limitations (Section 8) that "as video models become more and more capable of generating realistic, natural-looking videos, it is possible that future video models may not show the same statistical discrepancies between real and generated videos anymore." But this honest admission is not a weakness — it is a meta-innovation in how detection research communicates its boundaries. Rather than making an unbounded performance claim ("our method detects AI video"), ReStraV makes a conditional claim ("our method detects AI video as long as generative models fail to replicate the temporal smoothness that natural-image-trained representations expect").

This conditional framing has several important consequences:

It defines the research agenda for evading detection. A generative model developer who wants to defeat ReStraV knows exactly what to target: train the model to produce videos whose DINOv2 trajectories have curvature statistics matching natural video distributions. This could be achieved through adversarial training against the curvature metric, through architectural innovations that enforce temporal smoothness in representation space, or through training on data that includes the trajectory statistics as a quality signal. The arms race is specified in concrete, measurable terms rather than vague notions of "more realistic video."

It defines the research agenda for improving detection. If future generators do close the curvature gap, the detection framework doesn't become obsolete — it becomes the baseline against which more sophisticated geometric analysis is developed. Higher-order trajectory statistics (kurtosis of curvature, as the paper mentions in Section 8), multi-scale temporal analysis, or representations from different training distributions could capture residual discrepancies that simple mean curvature misses. The framework is extensible because it's grounded in a principle (natural temporal statistics vs. synthetic deviations) rather than a specific implementation.

It provides a criterion for evaluating detection claims. A future detection method that claims to outperform ReStraV can be evaluated not just on accuracy but on whether it captures a different geometric property — one that would remain even if curvature were matched. This prevents the field from converging on a single metric that, once optimized against, becomes useless (Goodhart's law, which the paper explicitly invokes in Section 8).

This innovation is meta-methodological rather than technical. It doesn't improve detection performance; it improves how detection claims are formulated and evaluated. By making the detection criterion explicit and falsifiable, ReStraV shifts the conversation from "does this method work?" (a question with a shelf life measured in months) to "under what conditions does this method work, and what would need to change for it to stop working?" (a question with lasting scientific value).

The evidence for the falsifiability of the hypothesis comes from the generalization experiments: Sora, a state-of-the-art generator not in the training set, shows a reduced but still substantial detection gap (80.05% accuracy, Table 3), while Veo3 — described as producing "videos with plausible physical interactions and consistent object interactions" — yields 83.2% accuracy in zero-shot evaluation. These numbers are lower than for older generators (where accuracy exceeds 90%), suggesting that as generators improve, the curvature gap may be narrowing — exactly the trend that the falsifiable hypothesis predicts. The paper doesn't have enough data points to establish a trend, but it provides the framework for measuring one.

5. Experimental Analysis

Evaluation Methodology

Dataset. The primary benchmark is VidProM (Wang & Yang, 2024), a million-scale dataset of real text-to-video prompts paired with AI-generated videos from multiple generators, using the 500 test questions from Lightman et al. (2022) for the core difficulty binning analysis. Additional evaluations use GenVidBench (Ni et al., 2025) for cross-generator generalization testing, Physics-IQ (Motamed et al., 2025) for matched real/generated video pairs depicting physical interactions, and the DeMamba benchmark (Chen et al., 2024) for one-to-many generalization assessment. For classifier training (Section 6), the paper constructs a balanced dataset of 50,000 AI-generated videos (10,000 each from Pika, VideoCrafter2, Text2Video-Zero, ModelScope, and Sora) sampled from VidProM, paired with 50,000 natural videos randomly selected from DVSC2023 (Pizzi et al., 2024). The test set for the main classifier evaluation mirrors this composition with a stratified 50/50 train/test split, maintaining identical class priors and per-generator balance in both subsets. For the benchmark comparisons against image-based detectors (Section 7A), the test set uses 40,000 real videos and 10,000 AI-generated videos from each of four models (Pika, VideoCrafter2, Text2Video-Zero, ModelScope), replicating the evaluation protocol from VidProM's original implementation.
Base model(s). The feature extractor is DINOv2 ViT-S/14 (Oquab et al., 2024), a self-supervised vision transformer with 22 million parameters pre-trained on a large and diverse collection of natural images using a student-teacher distillation objective. The model is used frozen — no fine-tuning, no adaptation, no exposure to video data during training. This choice is justified through a systematic comparison of 14 visual encoders (Section 4, Figure 2B) showing that DINOv2 produces the largest positive curvature gap (Δθ = 45.46°) between AI-generated and natural videos among all tested encoders, and that this gap is driven by differential straightening (natural videos straightened substantially, AI videos not) rather than absolute straightening capability. The ViT-S/14 variant is selected for its balance of representational quality and computational efficiency; the embeddings are extracted from the final transformer block (block.11) as the concatenation of the CLS token (384 dimensions) and all 196 patch tokens (384 dimensions each), yielding a 75,648-dimensional feature vector per frame. The classifier evaluated on top of these features is a two-layer MLP with architecture 64 → 32, trained with a 3-fold grid/random hyperparameter sweep and decision threshold optimized on the training set to maximize F1-score.
Metrics. The paper uses several standard detection metrics computed from the confusion matrix of the binary classification task (natural vs. AI-generated): accuracy (overall fraction of correct predictions), balanced accuracy (average of sensitivity and specificity, compensating for any class imbalance), specificity (true negative rate — fraction of natural videos correctly classified), precision (positive predictive value for the AI-generated class — fraction of AI predictions that are correct), recall (true positive rate for the AI-generated class — fraction of AI videos correctly detected), F1-score (harmonic mean of precision and recall for the AI-generated class), and AUROC (area under the receiver operating characteristic curve, measuring discriminative ability across all decision thresholds). For the Physics-IQ evaluation (Section 7E), the metric is two-alternative forced-choice (2AFC) accuracy — the fraction of trials where the model correctly identifies the AI-generated video from a matched real/generated pair, with chance performance at 50%. For the VidProM comparison against image-based detectors (Section 7A), mean Average Precision (mAP) is also reported, following the original VidProM evaluation protocol. Inference time is measured in milliseconds per video, computed as end-to-end latency (T_DINOv2 + T_clf) averaged over the test fold on a single NVIDIA RTX-2080 GPU. All performance numbers are reported on held-out test data, with classifier decision thresholds optimized on training data and applied unchanged to the test set.
Baselines. The paper compares ReStraV against a comprehensive set of existing detection methods across multiple evaluation settings. Image-based detectors (Section 7A, Table 2): CNNSpot (Chang et al., 2024) — a prompt-tuned vision-language model for fake image detection; FreDect (Doloriel & Cheung, 2024) — frequency masking for universal deepfake detection; Fusing (Wang et al., 2020) — CNN-generated image detection using classifier features; Gram-Net (Liu et al., 2020) — global texture enhancement for fake face detection; GIA (Frank et al., 2020) — frequency analysis for GAN-generated image recognition; LNP (Tan et al., 2023) — learning on gradients for GAN-generated image detection; DFD (Bi et al., 2023) — detecting generated images by real images only; UnivFD (Ojha et al., 2024) — universal fake image detection across generative models. Video-based detectors (Section 7B, Tables 3-5): VideoSwin (Liu et al., 2022) — video swin transformer adapted for detection; TSM (Lin et al., 2019) — temporal shift module for efficient video understanding; I3D (Carreira & Zisserman, 2017) — inflated 3D ConvNet for action recognition; SlowFast (Feichtenhofer et al., 2019) — two-pathway network for video recognition; X3D (Feichtenhofer, 2020) — expanded architectures for efficient video recognition; MVIT-V2 (Li et al., 2022) — improved multiscale vision transformers; TPN (Yang et al., 2020) — temporal pyramid network for action recognition; UniFormer-V2 (Li et al., 2022) — spatiotemporal learning by arming image ViTs with video UniFormer; TimeSformer (Bertasius et al., 2021) — space-time attention for video understanding; TALL (Xu et al., 2024) — thumbnail layout for deepfake video detection; NPR (Tan et al., 2024) — rethinking up-sampling operations for generalizable deepfake detection; STIL (Zhang et al., 2024) — spatiotemporal inconsistency learning for deepfake video detection; DeMamba (Chen et al., 2024) — AI-generated video detection on million-scale benchmark. VLM baseline (Section 7E, Figure 7): Gemini 1.5 Pro (Gemini Team et al., 2024) — a state-of-the-art multimodal large language model evaluated on the 2AFC Physics-IQ task, with results reported from Motamed et al. (2025). For the majority of comparisons, baseline results are taken directly from the respective benchmark papers (VidProM, GenVidBench, DeMamba) rather than re-implemented, ensuring fair comparison against published numbers.
Generation budget / compute accounting. The paper does not use "generations" as a compute budget in the traditional LLM sense, since DINOv2 is a frozen encoder with no sampling or search. Instead, computational cost is measured in inference time — the end-to-end wall-clock latency for processing a single video, measured in milliseconds on a single NVIDIA RTX-2080 GPU (8GB VRAM). The DINOv2 forward pass for an 8-frame batch takes 43.6 ms, with the total end-to-end latency for a 24-frame video (with the MLP classifier) being approximately 48 ms. This is the universal cost metric: all reported times include both feature extraction and classification. For the FLOPs-matched comparisons that characterize much of the LLM scaling literature, the paper does not perform an analogous analysis — there is no comparison of "how much pretraining compute equals how much inference compute," since the feature extractor is frozen and the only trained component is a tiny MLP. The efficiency comparison to baselines is qualitative rather than quantitative: ReStraV processes video at approximately 48 ms end-to-end, which is characterized as "a low-cost alternative to resource-intensive methods" (Section 6), but no FLOP counts, parameter counts, or training-time comparisons are provided for the competing video detectors. The computational simplicity is itself framed as a contribution.
Cross-validation / statistical protocol. For the main classifier training (Section 6), the paper uses a single stratified 50/50 train/test split with class priors identical in both subsets and balanced representation across the five AI generators (Pika, VideoCrafter2, Text2Video-Zero, ModelScope, Sora). The decision threshold for each classifier is optimized on the training set to maximize F1-score, then applied unchanged to the test set. For the ablation studies on video length and sampling density (Appendix B), multiple runs with different random video samples and 50/50 train-test partitioning are performed, with error bars in the figures representing ±1 standard deviation. The paper does not employ k-fold cross-validation for the main classifier evaluation, nor does it report confidence intervals or statistical significance tests on the benchmark comparisons (with the exception of the distributional analysis in Section 5, where t-tests and ANOVA are used to confirm feature separability). The difficulty-bin cross-validation protocol that characterizes the LLM inference-scaling paper (two-fold cross-validation within each difficulty quintile) has no direct analog here, since ReStraV does not involve difficulty estimation or adaptive strategy selection.

Main Quantitative Results

Classifier Performance on VidProM (Section 6, Table 1)

The central classifier evaluation tests six off-the-shelf models on the balanced 50,000/50,000 natural/AI-generated test set, with features computed from 24-frame DINOv2 trajectories. The headline result: the two-layer MLP (64 → 32) achieves 94.17% accuracy, 94.19% balanced accuracy, 98.88% precision, 94.14% recall, 96.48% F1-score, and 98.63% AUROC, with an end-to-end inference time of 48.12 ms per video (Table 1). This substantially outperforms linear classifiers — logistic regression achieves 89.02% accuracy and 95.26% AUROC, SVM with RBF kernel achieves 85.23% accuracy and 93.27% AUROC — confirming that the feature space is not linearly separable and that non-linear interactions between geometric statistics matter for optimal classification. Among the non-linear models, random forest achieves slightly higher recall (97.05%) but substantially lower specificity (80.37%), meaning it misclassifies approximately 19.6% of natural videos as AI-generated — a high false positive rate that would be problematic in deployment. Gradient boosting achieves the second-highest overall accuracy (92.83%) with strong precision (98.25%). The MLP's combination of high accuracy, balanced specificity (94.11%), and precision (98.88%) makes it the recommended configuration; its inference time is within approximately 4 ms of the fastest classifier (Gaussian Naive Bayes at 44.53 ms), with the DINOv2 forward pass (43.6 ms) dominating the total latency. The ROC curve for the MLP (Figure 9A) shows near-perfect separation, with the curve hugging the top-left corner across a wide range of thresholds, and the normalized confusion matrix (Figure 9B) reveals symmetric error rates: 6.7% false positive rate (natural videos misclassified as AI) and 5.5% false negative rate (AI videos misclassified as natural).

Comparison to Image-Based Detectors on VidProM (Section 7A, Table 2)

ReStraV is benchmarked against eight state-of-the-art image-based detectors on the VidProM dataset using a balanced test set of 40,000 real videos and 10,000 AI-generated videos from each of four generators (Pika, VideoCrafter2, Text2Video-Zero, ModelScope). Results are reported as per-generator accuracy and mean Average Precision. The baseline methods achieve moderate performance ranging from 45.32% (LNP) to 63.64% (Gram-Net) average accuracy, with substantial generator-dependent variation: Gram-Net achieves 84.19% on Pika but only 50.46% on ModelScope; FreDect ranges from 50.07% on Pika to 69.88–69.94% on the two Text2Video-based generators. ReStraV achieves 97.06% average accuracy (Pika: 90.90%, VideoCrafter2: 99.50%, Text2Video-Zero: 99.05%, ModelScope: 98.37%) and 98.81% average mAP (individual generator mAPs ranging from 98.44% to 99.12%). This represents a margin of approximately 33.4 percentage points over the best image-based baseline (Gram-Net at 63.64%). The paper explicitly acknowledges that "comparing ReStraV to image-based detectors on a video task is not an even comparison" (Section 7A), but the stark gap illustrates that frame-level spatial/frequency analysis — the operating regime of all eight baselines — fundamentally misses the temporal dimension where generative models currently fail most conspicuously. The per-generator pattern is informative: ReStraV's weakest performance is on Pika (90.90%), which is also where Gram-Net performs best (84.19%), suggesting Pika may have stronger frame-level artifacts but weaker temporal consistency, or vice versa — the two detection signals appear to capture different aspects of synthetic video.

Comparison to Video-Based Detectors on VidProM (Section 7B, Table 3)

The comparison against a dedicated video detector (VideoSwin Transformer) on VidProM is structured around three generalization scenarios: seen generators (training and testing on the same four generators), unseen generators (training on all four generators except VideoCrafter2 and Text2Video-Zero, which are held out for testing), and future generators (testing on Sora, which was not available during training). In the seen-generator condition, VideoSwin achieves 77.91% accuracy and 75.33% mAP, while ReStraV achieves 97.05% accuracy and 98.78% mAP — a margin of approximately 19 percentage points. In the unseen-generator condition, VideoSwin's performance degrades to 62.44% accuracy and 59.61% mAP, while ReStraV maintains 89.45% accuracy and 97.32% mAP. In the future-generator condition (Sora), VideoSwin drops further to 60.70% accuracy and 58.20% mAP, while ReStraV achieves 80.05% accuracy and 92.85% mAP. Two patterns emerge: first, ReStraV substantially outperforms VideoSwin across all conditions, and second, VideoSwin exhibits the classic generalization collapse (77.91% → 62.44% → 60.70%) characteristic of detectors that learn generator-specific artifacts, while ReStraV's degradation is more gradual (97.05% → 89.45% → 80.05%). The drop on Sora (80.05%) relative to seen generators (97.05%) suggests that as generative models improve, the curvature gap may narrow — Sora is a more advanced generator than the training models, and its trajectories appear to be somewhat closer to natural video statistics. However, a detection accuracy of 80.05% in zero-shot conditions is still practically useful and substantially exceeds the near-chance-level performance of VideoSwin.

Comparison to Video-Based Detectors on GenVidBench (Section 7B, Table 4)

The GenVidBench evaluation tests generalization across a substantially more diverse set of generators in two tasks: the Main (M) task (training on Pika, VideoCrafter2, Text2Video-Zero, ModelScope; testing on MuseV, Stable Video Diffusion, CogVideo, Mora, and natural videos from HD-VG/130M) and the Plants (P) task (same training generators, testing on plant-specific videos from the same test generators — described as "the most challenging subset" due to complex natural variation in flora). Results are reported as per-generator accuracy with the average across test generators. For the Main task, ReStraV achieves 93.01% average accuracy, with per-generator accuracies of 93.52% (MuseV), 94.01% (SVD), 93.52% (CogVideo), 92.97% (Mora), and 91.07% (HD-VG/130M natural videos). This substantially exceeds all baselines, which show both lower averages and extreme generator-dependent inconsistency: I3D achieves 60.11% on CogVideo but 8.15% on MuseV; SlowFast ranges from 12.25% to 93.63%. Even the best baseline performers show problematic variance — X3D achieves 92.39% on MuseV but only 37.27% on SVD, while MVIT-V2 achieves 98.29% on SVD but 47.50% on CogVideo. ReStraV's performance, by contrast, is notably stable across generators (range: 91.07%–94.01%, a spread of only 2.94 percentage points), suggesting the curvature signal is relatively generator-agnostic. For the Plants task, ReStraV achieves 93.96% average accuracy (MuseV: 95.06%, SVD: 97.83%, CogVideo: 92.38%, Mora: 91.24%, HD-VG/130M: 93.31%), again leading all baselines. The plants task is notable because it presents a domain where natural variation (irregular leaf patterns, wind motion, stochastic growth) might be expected to mask generative artifacts; the fact that ReStraV maintains high performance suggests the curvature signal captures temporal inconsistencies that are distinct from domain-specific natural variation. The boxplot visualization in Figure 13 (Appendix A.5) confirms that ReStraV's accuracy distribution is both higher and tighter than baselines across generators in both tasks.

One-to-Many Generalization on the DeMamba Benchmark (Section 7C, Table 5)

The one-to-many generalization test evaluates an extreme form of distribution shift: training on videos from a single generator (Pika, SEINE, or OpenSora) and testing on ten unseen generators (Sora, MorphStudio, Gen-2, HotShot, Lavie, Show-1, MoonValley, Crafter, ModelScope, WildScrape). Results are averaged across the three training conditions and compared against specialized deepfake video detectors (TALL, NPR, STIL, and DeMamba). Performance is reported as recall (R), F1-score, and Average Precision (AP). ReStraV achieves competitive or superior performance across all training conditions: when trained on Pika, ReStraV achieves F1 = 0.827 and AP = 0.797 (vs. DeMamba's F1 = 0.726, AP = 0.817); when trained on SEINE, ReStraV achieves F1 = 0.898 and AP = 0.854 (vs. DeMamba's F1 = 0.787, AP = 0.894); when trained on OpenSora, ReStraV achieves F1 = 0.797 and AP = 0.717 (vs. DeMamba's F1 = 0.671, AP = 0.738). ReStraV achieves the highest F1-score in all three training conditions, while DeMamba achieves the highest AP in two of three. This is notable because DeMamba is a dedicated video detection architecture with millions of parameters trained specifically for this task, while ReStraV uses frozen DINOv2 features and a tiny MLP. The strong performance when training on only a single generator — arguably the most extreme generalization test in the paper — provides evidence that the curvature signal is not merely providing incremental improvement over existing detectors but capturing a fundamentally different, more generalizable property of synthetic video.

Zero-Shot Generalization to Veo3 (Section 7D)

The zero-shot generalization test evaluates ReStraV on Google's Veo3, a state-of-the-art model "acclaimed for its ability to generate videos with plausible physical interactions and consistent object interactions" and explicitly noted as not included in any training data. Using only the MLP classifier trained in Section 6 (on Pika, VideoCrafter2, Text2Video-Zero, ModelScope, and Sora — no Veo3 videos), ReStraV is tested on 200 Veo3 versus 200 natural video pairs and achieves 83.2% accuracy, 85.1% F1-score, and 86.9% AUROC. This is lower than performance on seen generators (~94–99% depending on the specific model) but substantially above chance (50%), confirming that Veo3 — despite its advanced temporal consistency — still produces representation-space trajectories that are distinguishable from natural videos by the ReStraV signal. The 83.2% accuracy on Veo3 is comparable to the 80.05% achieved on Sora in the future-generator condition (Table 3), suggesting these two advanced generators produce similar degrees of curvature anomaly. The paper does not break down the Veo3 results by video content type or difficulty, so it is unclear whether the misclassified ~17% of Veo3 videos represent a hard subset where the model achieves near-natural temporal smoothness, or whether the errors are randomly distributed.

Comparison to VLM on Physics-IQ (Section 7E, Figure 7)

The Physics-IQ evaluation tests detection on a uniquely challenging dataset: matched pairs of real and AI-generated videos where both videos in each pair depict the same physical scene (identical starting frame, identical objects, identical lighting conditions), with the only difference being that one video is real footage and the other is generated by Runway, Pika, or Sora. This controlled setting removes confounds like different content or scene composition that could provide trivial detection cues, forcing the detector to rely purely on the temporal dynamics of the generated content. The evaluation uses a two-alternative forced-choice (2AFC) paradigm: for each pair, the method predicts which video is AI-generated by selecting the one with higher mean curvature (no classifier training, no calibration — raw curvature comparison). Results are compared against Gemini 1.5 Pro, a state-of-the-art vision-language model evaluated by Motamed et al. (2025) on the same task. Gemini achieves 74.8% on Runway videos, 80.5% on Pika videos, and only 55.6% on Sora videos — near chance for the most photorealistic generator. ReStraV achieves 97.5% for Pika, 94.9% for Runway, and 99.0% for Sora (Figure 7). The near-ceiling performance on Sora is particularly striking given Gemini's near-chance performance: Sora's photorealism, which makes it difficult for even a sophisticated VLM to distinguish from real footage, does not translate to naturalistic curvature trajectories in DINOv2 representation space. This suggests that visual photorealism and representational temporal smoothness are dissociable properties — a video can look perfect to human (and VLM) perception while still exhibiting geometric trajectory anomalies that ReStraV detects. The fact that the detection is done with a simple rule (higher mean curvature = AI-generated) rather than a trained classifier underscores the robustness of the curvature signal in this controlled setting.

Distributional Analysis of Geometric Features (Section 5, Figure 5)

Beyond classifier evaluation, the paper presents a large-scale distributional analysis comparing 50,000 AI-generated and 50,000 natural videos along the eight aggregated statistical features (mean, min, max, variance of both distance and curvature). Figure 5 shows clear distributional separation for all eight features, with the curvature-based features exhibiting the strongest discriminative signal: the mean curvature (μ_θ) distribution for AI-generated videos is shifted substantially rightward (higher values), indicating less straightening, while the curvature variance (σ_θ^2) distribution shows natural videos with higher variance — the "low mean, high variance" pattern the paper discusses in Section 8. Distance-based features also show separation but with more overlap. Statistical tests confirm these observations: a two-sample t-test comparing μ_θ yields t = -44.02, p ≈ 0; an ANOVA across AI generators and natural videos yields F = 18,598.17, p ≈ 0. These extremely strong statistical separations demonstrate that the geometric features provide a reliable discriminative signal even before any classifier is trained. The per-generator curvature trajectories in Figure 8 (Appendix A.1) provide qualitative examples: natural videos (blue) typically show smoother, lower-amplitude curvature trajectories, while AI-generated videos (red) from Sora, Pika, and Runway exhibit more erratic fluctuations with higher peaks — though the paper notes that "individual trajectories can be noisy."

Feature Importance Analysis (Appendix A.3, Figure 11)

Permutation feature importance analysis on the MLP classifier reveals that mean curvature (μ_θ) is unequivocally the most critical feature, with importance "more than double that of the next most influential feature" (curvature variance σ_θ^2). The eight aggregated statistical features occupy the top importance ranks, with individual time-step signals (d_i, θ_i) providing "complementary but less critical information." This validates the core hypothesis: the overall straightness of the trajectory — not specific frame-to-frame transitions — is the primary discriminative signal, and the remaining 13 features (time-step values) contribute incrementally at best. The implication is that a detector using only the 8 aggregated statistics would likely approach the full 21-dimensional classifier's performance, though the paper does not present this ablation.

Ablation Studies and Robustness Checks

Video length and sampling density (Appendix B.1, Figure 15): The paper systematically varies the analyzed video length (1–5 seconds, at 30 FPS with every-3rd-frame sampling) and sampling density (within a fixed 2-second window, varying the frame sampling interval k). For video length, performance improves monotonically: AUROC increases from approximately 92% at 1 second (T ≈ 10 frames) to approximately 98% at 5 seconds (T = 50 frames), with inference time increasing linearly. The 2-second, 24-frame configuration achieves AUROC exceeding 96% and is selected as the accuracy-efficiency trade-off. For sampling density within a 2-second window (60 total frames at 30 FPS), the optimal interval is k = 3 (T = 20 frames, AUROC ≈ 97%), with degradation at both sparser (k = 5, T = 12) and denser (k = 1, T = 60) sampling. The paper notes that the degradation at very dense sampling (every frame, T = 60) is counterintuitive but consistent with the idea that excessively fine-grained temporal resolution introduces noise without improving the geometric signal — adjacent frames at 30 FPS are highly redundant, and the curvature computation benefits from the larger representational changes that occur over longer inter-frame intervals.
Temporal window position (Appendix B.2, Figure 16): Sliding a 2-second window (T = 24 frames) across 5-second videos with a 10-frame step reveals that detection performance is robust to where in the video the analysis is applied. Performance metrics (accuracy, AUROC, F1) remain within a narrow band (approximately 95–97% AUROC), with a slight U-shaped trend — marginally higher at the beginning and end of the analyzed 0–3 second start-time range, marginally lower in the center. The robustness is important for practical deployment: the system does not require temporal alignment, shot boundary detection, or intelligent window selection.
Scene cut frequency as a potential confound (Appendix C, Figure 17): The paper addresses the concern that higher scene cut frequency in AI-generated videos could drive the curvature signal by analyzing 13,000 AI-generated and 13,000 natural videos using PySceneDetect. Results show that the average number of hard scene cuts in the first 2 seconds is low and comparable across sources: most AI generators (Pika, Show-1, ModelScope, Gen-2, HotShot, Crafter, MoonValley, Text2Video-Zero) show 0.00 cuts on average, while natural videos average 0.08 cuts. The highest AI generator value is VideoCrafter2 at 0.10 cuts — comparable to natural video rates. The paper concludes that scene cut frequency is "not a significant confounding factor," particularly given the short 2-second analysis window.
Encoder selection: differential vs. absolute straightening (Section 4, Figure 2B): The comparison of 14 visual encoders reveals that detection-relevant curvature gap (Δθ) is uncorrelated with absolute straightening capability (ρ = -0.13, p = 0.64). The best absolute straighteners — HVS-inspired models (Gabor filter bank, LGN-V1) — produce negative curvature gaps (natural videos more curved than AI videos), making them useless for detection. DINOv2 produces the largest positive gap (45.46°), substantially exceeding other self-supervised models (CLIP, BYOL, SimCLR), supervised CNNs, and spatio-temporal models. This ablation justifies the choice of DINOv2 as the feature extractor and establishes that detection performance depends on selective (not absolute) straightening.
Classifier comparison (Section 6, Table 1, Figure 6): Six off-the-shelf classifiers are evaluated on identical 21-dimensional features. The MLP achieves the best overall performance (94.17% accuracy, 98.63% AUROC), followed by random forest (94.24% accuracy, 98.03% AUROC) and gradient boosting (92.83% accuracy, 97.85% AUROC). Linear models (logistic regression, SVM) underperform, confirming non-linear feature interactions. The Gaussian Naive Bayes classifier achieves the fastest inference (44.53 ms) but substantially lower accuracy (86.64%). Figure 6 visualizes the speed-accuracy trade-off, with the MLP occupying the upper-left region (high accuracy, low latency), making it the recommended configuration. The MLP's confusion matrix (Figure 9B) shows symmetric error rates (6.7% false positive, 5.5% false negative), indicating no strong class bias.
Raw curvature trajectory examples (Appendix A.1, Figure 8): Qualitative examples of curvature trajectories for matched natural/AI-generated video pairs from Sora, Pika, and Runway illustrate the variability in individual trajectories. While aggregate statistics separate cleanly (Figure 5), individual videos show noisy curvature sequences; the AI-generated trajectories tend to display "different overall levels of curvature, more pronounced peaks, or more erratic behavior compared to the often smoother or distinctly patterned trajectories of natural videos." This underscores why statistical aggregation is necessary — the raw curvature sequences are too noisy for simple thresholding, but their distributional properties provide reliable separation.
Decision boundary visualization (Appendix A.2, Figure 10): Voronoi tessellations of the 2D projection of the feature space for logistic regression and MLP classifiers illustrate the nature of the learned decision boundaries. The LR boundary is a single linear hyperplane, while the MLP learns a more complex, non-linear boundary that better separates the interleaved natural and AI-generated clusters. This visualization confirms that the feature space is not linearly separable and that the MLP's additional capacity is needed for optimal discrimination.
Feature importance stability (Appendix A.3, Figure 11): The permutation feature importance analysis with error bars (standard deviation across permutations) confirms that the dominance of aggregated statistics — particularly μ_θ — is robust to permutation sampling. The error bars are small relative to the importance magnitudes, indicating consistent feature rankings across different random permutations. Individual time-step signals have small importance values with overlapping error bars, suggesting that no single time step carries unique discriminative information beyond what the aggregated statistics capture.
Negative result: ReST^EM fine-tuning degrades revision models (Appendix K, Figure 16): This is not directly applicable to ReStraV (which uses a frozen encoder) but the paper includes extensive training-data construction analysis for the video domain. No analogous negative result is reported for the ReStraV pipeline — the paper does not present experiments where alternative feature extractors, curvature definitions, or aggregation strategies were tried and found to underperform, with the exception of the encoder comparison (which did produce a negative result: HVS-inspired models and spatio-temporal models are unsuitable).

Critical Assessment

Does ReStraV Demonstrate That the Perceptual Straightening Hypothesis Enables Detection?

The paper's central narrative is that detection is enabled by "perceptual straightening" — the neuroscientific finding that biological and artificial visual systems straighten natural video trajectories. However, a careful reading of the evidence reveals that what actually enables detection is differential straightening — the selectivity with which DINOv2 straightens natural versus synthetic trajectories — which is conceptually distinct from, and empirically uncorrelated with, perceptual straightening as studied in neuroscience. The paper acknowledges this explicitly: "Our findings indicate that the best perceptual straighteners do not correspond to the best fake video detectors" (Section 4), and the correlation between absolute straightening and the curvature gap is ρ = -0.13, p = 0.64 — essentially zero. The HVS-inspired models, which are the closest computational instantiations of the biological straightening mechanism, are the worst detectors because they straighten everything equally.

This means the paper's title ("AI-Generated Video Detection via Perceptual Straightening") and much of its framing overstate the connection. The method works not because perceptual straightening occurs — it works because DINOv2's self-supervised training on natural images produces representations that are selectively tuned to natural statistics, and synthetic videos violate those statistics. The connection to perceptual straightening is inspirational rather than mechanistic: the idea of measuring trajectory curvature comes from the straightening literature, but the phenomenon that produces the detection signal (differential straightening) is a distinct empirical finding that was not predicted by and does not follow from the perceptual straightening hypothesis. The paper would be more accurately titled "AI-Generated Video Detection via Differential Representational Straightening in Self-Supervised Vision Transformers" — but this nuance is buried in Section 4 and the Discussion, while the framing (abstract, introduction, Figure 1) emphasizes the neuroscience connection.

This does not invalidate the method — ReStraV works, and works well — but it means the theoretical contribution is somewhat different from what is claimed. The genuine theoretical contribution is the discovery that frozen self-supervised image encoders can serve as temporal anomaly detectors without any video training, and that the relevant property of the encoder is not how well it straightens trajectories absolutely but how selectively it straightens them. This is an important finding, but it is not a direct application of perceptual straightening — it is a repurposing of a measurement technique (curvature) from the straightening literature into a detection context, combined with an empirical discovery about which models produce useful curvature signals.

Does ReStraV's Performance Genuinely Reflect Temporal Anomaly Detection, or Could Simpler Signals Suffice?

The paper's framing emphasizes that ReStraV captures "temporal inconsistencies" and "geometric distortions in the temporal trajectory of neural representations." However, the method operates on frame-level embeddings extracted independently by an image encoder — DINOv2 processes each frame in isolation. The temporal information is captured entirely by the sequence of these independent embeddings, not by any mechanism in the feature extractor that models time. This raises a question: is the detection signal genuinely temporal, or could it be reduced to frame-level differences that don't require sequential analysis?

Consider what the curvature measure captures: the angle between successive displacement vectors in representation space. If AI-generated videos produce frames that are simply more variable from frame to frame (higher stepwise distance d_i) — perhaps due to inconsistent textures, flickering objects, or unstable generative processes — this alone could produce higher curvature, since curvature depends on both the magnitudes and directions of successive displacements. The paper's distributional analysis (Figure 5) shows that AI-generated videos have higher mean distance (μ_d) and higher distance variance (σ_d^2) — they change more, and more erratically, from frame to frame. If the increased curvature is largely driven by increased frame-to-frame variability, then the temporal trajectory analysis may be adding relatively little beyond what a simpler frame-difference analysis would capture.

The paper does not present an ablation that isolates curvature from distance effects. Specifically, it does not test whether a classifier trained only on distance features (the 8 distance statistics) performs comparably to one trained on both distance and curvature, or whether curvature provides unique information beyond what distance captures. The feature importance analysis (Figure 11) shows that μ_θ is the single most important feature, but this doesn't establish that curvature carries information independent of distance — the two are correlated by definition (curvature is computed from normalized displacement vectors).

This is not to say the temporal analysis is unnecessary. The Physics-IQ results (Figure 7) are particularly suggestive: on a dataset where real and generated videos have identical starting frames, the curvature signal achieves near-perfect detection, suggesting that the temporal trajectory shape genuinely differs. And the generalization results (Tables 3–5) show robustness that frame-level artifact detectors lack. But the paper does not rigorously establish that the temporal geometric signal (curvature as opposed to frame-to-frame magnitude changes) is the active ingredient, as opposed to frame-to-frame variability more broadly. An ablation comparing ReStraV against a classifier trained on frame-level statistics without sequential ordering (e.g., bag-of-frames variance) would clarify this, but is not presented.

How Meaningful Is the 4× Efficiency Claim?

The paper does not make a 4× efficiency claim in the style of the LLM inference-scaling paper; instead, it emphasizes that ReStraV is "computationally efficient" and processes video in "approximately 48 ms end-to-end." This is compared qualitatively to "resource-intensive methods" without providing specific latency or FLOP measurements for the competing video detectors (VideoSwin, TimeSformer, etc.). The efficiency claim thus rests on a qualitative assertion rather than a quantitative comparison. A reader wanting to know whether ReStraV is faster than VideoSwin by a factor of 2, 10, or 100 cannot determine this from the paper — no baseline latency measurements are provided.

The 48 ms figure itself requires scrutiny. The paper reports that the DINOv2 forward pass takes 43.6 ms for an 8-frame batch (Appendix D, Section 6), and that the total end-to-end latency is approximately 48 ms. For a 24-frame video, this implies either that frames are processed in larger batches than 8 (e.g., all 24 at once, which might fit in GPU memory but is not specified), or that the 48 ms includes only the classifier time with the DINOv2 cost amortized differently. The computational budget is not precisely specified, making it difficult to replicate or fairly compare against other methods.

Do the Benchmark Comparisons Actually Demonstrate State-of-the-Art Performance?

The paper claims that ReStraV "improve[s] detection accuracy that often surpasses state-of-the-art methods" (contributions list, Section 1) and "set[s] a new SoTA in fake video identification" (Section 8). The evidence for this claim is strong in the settings tested: ReStraV substantially outperforms all baselines on VidProM (Tables 2, 3), GenVidBench (Table 4), Physics-IQ (Figure 7), and the DeMamba one-to-many benchmark (Table 5). The margins are large and consistent across evaluation settings.

However, several caveats should be noted. First, the baseline numbers for video detectors on VidProM and GenVidBench are taken from the respective benchmark papers rather than re-implemented and re-evaluated under identical conditions. This is standard practice, but it means that differences in preprocessing, frame sampling, or hardware could contribute to the performance gap. ReStraV uses 24 frames at 224×224 resolution extracted over 2 seconds; if the baselines were originally evaluated with different temporal sampling (fewer frames, different frame rates, different spatial resolutions), the comparison may not be entirely fair. The paper does not discuss whether the baseline numbers reflect optimal configurations or whether hyperparameter tuning could close some of the gap.

Second, the "state-of-the-art" claim is made against a specific set of published methods. The rapid pace of AI video detection research means that by the time of publication, new methods may have appeared that narrow or close the performance gap. This is not a criticism of the paper — it's inherent to benchmarking fast-moving fields — but the "SoTA" claim should be understood as time-stamped to the baselines evaluated.

Third, the GenVidBench Plants task (Table 4b) shows that ReStraV's advantage is smaller against some baselines on some generators. TimeSformer achieves 96.30% on CogVideo in the Plants task while ReStraV achieves 92.38% — a reversal of the typical pattern. And on the natural video set (HD-VG/130M), ReStraV's 93.31% is lower than VideoSwin's 98.76%. This suggests that ReStraV's advantage is not uniform across all content types and generators, and that some video-specific architectures may capture signals that ReStraV misses on certain data subsets. The paper does not analyze these reversals or discuss what properties of plants videos or specific generators might favor different detection approaches.

The Difficulty Estimation Problem Is Avoided, Not Solved

The LLM inference-scaling paper's central challenge was estimating prompt difficulty to enable adaptive strategy selection — and that challenge was only partially solved (the PRM-based method requires 2048 samples per question). ReStraV sidesteps this problem entirely by applying a uniform analysis (same 24 frames, same DINOv2 encoder, same curvature computation) to all videos regardless of content or difficulty. This is both a strength (simplicity, no meta-decision required) and a potential weakness: the method might benefit from adaptive frame sampling (more frames for difficult videos, fewer for easy ones) or from content-aware window placement, but no such adaptation is explored. The robustness of performance across generators and content types (Tables 3–5) suggests that uniform analysis may be sufficient, but the paper does not investigate whether certain video types — long static shots, rapid action, abstract content — challenge the method in ways that adaptive analysis could address.

What Experiments Would Have Strengthened the Paper?

Several experiments could have addressed the gaps identified above:

Ablation of curvature versus distance features. Train classifiers on (a) only the 8 distance statistics, (b) only the 8 curvature statistics, and (c) both, to quantify the unique contribution of curvature beyond what inter-frame magnitude changes alone provide. This would directly test whether the geometric trajectory analysis adds value beyond simpler frame-difference measures.
Latency and FLOP comparisons to video detector baselines. Provide wall-clock inference time and/or FLOP counts for a representative video detector (e.g., VideoSwin) processing the same videos, enabling quantitative efficiency comparisons rather than qualitative claims.
Per-content-type breakdown. Report ReStraV's performance stratified by video content type (e.g., static scenes, dynamic action, talking heads, nature footage, abstract/synthetic-looking natural video) to identify failure modes and characterize the method's operating envelope.
Adversarial robustness evaluation. Test whether generative models can be fine-tuned to reduce ReStraV's curvature signal — for example, by adding a curvature-minimization loss during generation. This would test the falsifiable hypothesis that the paper articulates in Section 8.
Analysis of misclassified videos. Present and analyze examples of false positives (natural videos classified as AI) and false negatives (AI videos classified as natural) to understand what types of content confound the curvature signal. The paper provides qualitative trajectory examples (Figure 8) but does not analyze classifier errors.
Multi-encoder ensemble. Since the encoder analysis (Figure 2B) shows that different encoders produce different curvature gaps, combining signals from multiple encoders might improve robustness or capture complementary anomalies. The paper does not explore this, despite the low computational cost of running multiple frozen encoders.

Summary Assessment

The experiments provide strong evidence that trajectory curvature in DINOv2 representation space is a reliable signal for detecting AI-generated video, with ReStraV substantially outperforming both image-based and video-based detectors across multiple benchmarks, including challenging generalization settings (unseen generators, future generators, zero-shot evaluation). The performance margins are large and consistent, and the method's simplicity (frozen encoder, lightweight classifier, no video-specific training) makes it practically attractive. The ablation studies on video length, sampling density, and window position are thorough and justify the chosen operating point.

However, the experiments do not fully validate the theoretical framing (perceptual straightening as the mechanism) — the evidence more precisely supports differential straightening as the operative phenomenon, which is conceptually distinct. The paper also does not rigorously establish that the geometric trajectory analysis (curvature specifically, as opposed to simpler frame-to-frame variability measures) is necessary for the observed performance, nor does it provide the quantitative efficiency comparisons that would substantiate the "low-cost" claim. The benchmark comparisons, while strong, rely on published baseline numbers rather than controlled re-evaluation, and the paper does not analyze failure modes or content-type dependence in depth.

The generalization results (Tables 3–5, Veo3 zero-shot) are the most compelling evidence for the method's robustness, particularly the fact that ReStraV trained on a single generator can achieve competitive performance against dedicated video detectors on ten unseen generators (Table 5). This pattern — strong within-distribution performance combined with unusually good out-of-distribution generalization — is consistent with the paper's central claim: that curvature captures a fundamental geometric property of synthetic video trajectories rather than generator-specific artifacts. The Physics-IQ results (Figure 7) provide the cleanest demonstration of this signal, since the matched-pair design controls for content and isolates temporal dynamics.

The paper's most significant limitation is not empirical but conceptual: the connection to perceptual straightening is inspirational rather than mechanistic, and the framing somewhat obscures the genuine contribution — the discovery that frozen self-supervised image encoders can serve as effective temporal anomaly detectors through simple geometric analysis of their representational trajectories. This is an important finding that the experiments support robustly, even if the theoretical narrative requires some qualification.

6. Limitations and Trade-offs

Difficulty Estimation for Adaptive Allocation Is Not Solved — It Is Avoided

The paper proposes a uniform analysis pipeline: every video, regardless of content, duration, or generative source, is processed identically — 24 frames over 2 seconds, same DINOv2 encoder, same curvature and distance computation, same aggregated statistics, same MLP classifier. No mechanism adapts the sampling density, window placement, or feature computation based on video properties. While this simplicity is a strength in terms of deployment, it also means that the method cannot allocate more test-time compute to harder cases or less to easier ones — a capability that the compute-optimal scaling literature has shown can yield 4× efficiency gains in related detection domains.

The consequence is that ReStraV's computational budget is fixed per video regardless of whether the detection task is trivially easy (an obviously synthetic video from a weak generator) or extremely challenging (a near-photorealistic video from an advanced model like Veo3). The paper provides no framework for estimating, before or during processing, whether a given video is likely to be easily classifiable or near the decision boundary. A practical deployment might waste computation on videos that could be classified from 4 frames while providing insufficient signal for borderline cases that might benefit from longer temporal windows or multi-region analysis. The ablation studies (Appendix B.1, Figure 15) show that performance improves with longer videos (AUROC from ~92% at 1 second to ~98% at 5 seconds), confirming that additional temporal context is valuable — but the method has no way to selectively allocate that context where it would matter most.

The paper does not present a difficulty estimation mechanism, does not measure performance stratified by video "difficulty" (beyond per-generator breakdowns, which capture generator-specific rather than instance-specific variation), and does not explore whether the curvature signal itself could serve as an online difficulty indicator — for instance, high variance in early-frame curvature might predict classification uncertainty and trigger extended analysis. The authors do not acknowledge this as a limitation; the uniform pipeline is presented as a feature (simplicity, low cost) rather than as a constraint on adaptive allocation. The discussion of limitations (Section 8) focuses on the detection-generation arms race and the invariance to time reversal, not on the absence of difficulty-aware computation.

Mitigation status: None. The paper does not address adaptive allocation, does not propose difficulty estimation, and does not frame the fixed-budget uniform pipeline as a limitation. A practitioner deploying ReStraV in a cost-sensitive setting (e.g., large-scale batch processing of millions of videos) would need to determine whether the fixed per-video cost is acceptable or whether a two-stage approach — fast screening followed by deeper analysis for ambiguous cases — would improve efficiency. The paper provides no guidance on this trade-off.

The Detection Signal Depends on DINOv2's Specific Training Distribution — Generalization to Radically Different Domains Is Unproven

All of ReStraV's evaluations use natural videos as the reference class and AI-generated videos from diffusion- or transformer-based generative models as the target class. The natural video datasets (DVSC2023, HD-VG/130M, Physics-IQ real footage) represent conventional real-world content — human activities, natural scenes, everyday objects in motion. The detection signal relies on DINOv2 having internalized the temporal statistics of this type of content during its self-supervised pre-training on natural images, such that frame-to-frame transitions in natural videos respect the learned representational geometry and AI-generated transitions violate it.

What happens when the "natural" reference class shifts to content that DINOv2 has not been extensively trained on? Consider medical imaging videos (endoscopy, ultrasound), scientific visualizations (particle collision recordings, microscopy time-lapses), or artistic content (hand-drawn animation, stop-motion). These domains have temporal statistics that may differ substantially from the photographic natural images in DINOv2's training distribution. The paper provides no evidence that the differential straightening signal would persist — it is entirely possible that in these domains, natural videos would appear as "anomalous" to DINOv2 as AI-generated videos do in conventional domains, producing high curvature and false positives. Conversely, an AI generator specifically trained on medical imaging data might produce videos whose temporal statistics better match DINOv2's (limited) exposure to that domain, reducing the curvature gap and producing false negatives.

The encoder selection analysis (Section 4, Figure 2B) demonstrates that different visual encoders produce dramatically different curvature gaps — DINOv2 shows Δθ = 45.46° while spatio-temporal models show gaps of only 5–20° and HVS-inspired models show negative gaps. This sensitivity to encoder choice implies sensitivity to training distribution: if DINOv2's natural-image-trained representations are what create the gap, then videos from domains outside that training distribution may not benefit from the same differential straightening. The paper acknowledges that "biases inherited from pre-trained encoders (e.g., DINOv2) may cause fairness issues across diverse content" (Section 8, Broader Impacts), but frames this as a fairness concern rather than a fundamental limitation on domain generalization — and presents no cross-domain evaluation.

The consequence for practitioners is uncertainty about deployability in specialized domains. A forensic lab analyzing surveillance footage, a medical institution verifying diagnostic video authenticity, or a media organization screening user-generated content across diverse genres cannot assume that ReStraV's VidProM benchmark performance (97.06% accuracy) will transfer. The performance could be substantially lower, or the false positive rate could be unacceptably high, in domains where natural video statistics deviate from DINOv2's training manifold.

Mitigation status: The paper acknowledges the risk of encoder-inherited bias in the Broader Impacts section and calls for "rigorous bias auditing and debiasing efforts," but provides no empirical characterization of domain dependence and no systematic evaluation beyond the specific natural-video datasets used in the benchmarks. The encoder comparison (Figure 2B) provides indirect evidence that domain shift could be problematic (since encoder choice dramatically affects the signal), but does not test domain shift directly. A practitioner would need to conduct their own domain-specific validation before deploying ReStraV outside conventional natural video domains.

Curvature Is Invariant to the Arrow of Time — Physically Impossible Events That Are Not AI-Generated Will Be Missed

The paper explicitly notes: "Our method is invariant to playing a video backwards. This is clearly unnatural, if things, e.g., fall up instead of down; though at the same time this also would not be an instance of an AI-generated video" (Section 8, Discussion). The authors frame this as a limitation of their metric but also as a distinction between their detection target (AI-generated content) and physically implausible content more broadly.

However, this invariance constitutes a fundamental constraint on what types of synthetic manipulation the method can detect. Consider a video that has been manipulated by reversing a segment — a ball appearing to jump upward from the ground into a person's hand, smoke flowing backward into a chimney. This video would not be "AI-generated" in the sense of being produced from scratch by a generative model, but it would be synthetic, manipulated, and deceptive — exactly the kind of content that detection systems are expected to flag. ReStraV would see no curvature anomaly because the reversed trajectory, when played forward through the representation space, has the same curvature as the original forward trajectory (curvature is a symmetric function of the displacement vectors — reversing the sequence produces the same set of angles, just in reverse order). The distance values would also be identical, just reversed. The 8 aggregated statistics (mean, variance, min, max of curvature and distance) would be unchanged.

More broadly, any temporal manipulation that preserves local frame-to-frame relationships — reversal, uniform slowdown or speedup, frame duplication, or segment reordering that maintains smooth transitions — would leave the curvature and distance statistics largely intact. The method is specifically sensitive to frame-to-frame inconsistencies in representational geometry, not to global temporal coherence or physical plausibility. A video of a physically impossible event (a茶杯 floating upward) that was generated with perfectly smooth frame transitions would have low curvature and might be classified as natural, while a real video with a sudden lighting change or camera shake might have high curvature and be flagged as synthetic.

The paper acknowledges the time-reversal invariance but does not explore its scope: how many types of non-AI-generated manipulation would also evade detection? The Physics-IQ evaluation (Section 7E, Figure 7) tests detection of AI-generated videos of physical events, not detection of physically manipulated real videos. The distinction between "AI-generated" and "manipulated" is not merely academic — in practice, detection systems need to handle both, and ReStraV's signal is provably blind to a class of manipulations (temporal reversals and other symmetry-preserving edits) that are straightforward to perform with video editing software.

Mitigation status: The limitation is acknowledged in the Discussion, but no mitigation is proposed beyond the implicit suggestion that the method targets a specific threat model (fully AI-generated video) and that other methods would be needed for temporal manipulation detection. The paper does not evaluate whether combining curvature with a directional metric (e.g., measuring the arrow of time in learned representations, as cited in references [85, 86]) could recover sensitivity to temporal reversals while maintaining the benefits of the curvature signal. A practitioner concerned with manipulated content rather than purely generated content would need to supplement ReStraV with complementary detection mechanisms that are sensitive to temporal direction and physical plausibility.

The Headline Performance Numbers Do Not Include the Cost of Difficulty Estimation, and the Fixed-Budget Approach May Be Suboptimal for Large-Scale Deployment

Section 5 already noted this issue in the context of the experimental analysis, but it merits direct treatment as a limitation: the paper's reported accuracy and AUROC numbers assume that every video receives exactly the same computational budget (24 frames, 2-second window, DINOv2 forward pass, MLP classification), and no attempt is made to determine whether this fixed budget is optimal, or whether a variable-budget approach could achieve equivalent accuracy at lower average cost. The 48 ms per video figure is presented as evidence of efficiency, but it reflects the cost of a one-size-fits-all pipeline that processes every video identically regardless of how easy or hard the classification turns out to be.

This matters because real-world deployment often involves highly skewed difficulty distributions. A social media platform screening billions of user-uploaded videos daily might find that the vast majority are trivially natural (unmanipulated footage from smartphone cameras) or trivially synthetic (outputs from well-known generative models with distinctive artifacts), with only a small fraction occupying the ambiguous boundary where full 24-frame analysis is actually needed. Spending 48 ms on every video in such a setting means 48 ms × billions = enormous aggregate compute cost, the vast majority of which is wasted on easy cases that could be classified from far fewer frames. The ablation studies (Appendix B.1, Figure 15) show that even 1-second windows (T ≈ 10 frames) achieve AUROC around 92% — substantially above chance — suggesting that a two-stage pipeline (fast screening with a few frames, followed by full analysis only for ambiguous cases) could dramatically reduce average cost while maintaining accuracy.

The paper provides no framework for early-exit classification, no analysis of how classifier confidence relates to frame count, and no investigation of whether the curvature signal stabilizes after a small number of frames (enabling early termination when the trajectory is clearly natural or clearly synthetic). The classifier is trained on fixed 24-frame statistics; there is no evidence that it would perform well — or could be calibrated — for variable-length inputs. A practitioner wanting to minimize per-video cost in a high-volume setting would need to develop their own adaptive sampling strategy without guidance from the paper on how curvature statistics evolve with frame count or how to set early-stopping criteria.

The comparison to the compute-optimal scaling literature is instructive: that work showed that uniform budget allocation (e.g., best-of-256 for every prompt) leaves 4× efficiency gains on the table compared to difficulty-aware allocation. ReStraV's uniform 24-frame pipeline may be leaving similar efficiency gains unrealized, but the paper does not frame this as a limitation or explore adaptive alternatives. The absence is particularly notable given that the method's simplicity — curvature is computed from the same frozen encoder regardless of frame count — would make variable-length analysis straightforward to implement if appropriate stopping criteria could be defined.

Mitigation status: Not addressed. The paper does not discuss the trade-off between per-video cost and classification difficulty, does not propose adaptive sampling, and does not evaluate whether the reported accuracy could be maintained at lower average computational cost through selective frame allocation. The ablation studies on video length (Appendix B.1) show that longer windows improve accuracy but at linearly increasing cost, providing the raw data from which an adaptive strategy could be constructed — but the paper does not take this step.

The Method Has Not Been Evaluated Against Adversarial Attacks, and the Falsifiable Hypothesis Implies a Clear Attack Vector

The paper is admirably explicit about the conditions under which ReStraV would fail: "as video models become more and more capable of generating realistic, natural-looking videos, it is possible that future video models may not show the same statistical discrepancies between real and generated videos anymore" (Section 8, Limitations). It also invokes Goodhart's law: "when a measure becomes a target, it ceases to be a good measure," and acknowledges that "it is conceivable that someone developing a video model could train it in a way that optimizes for deceiving detection measures" (Section 8).

What the paper does not do is test whether current generative models can be adversarially adapted to reduce the curvature signal. This is a significant omission because the attack vector is straightforward and clearly specified by the method itself: if the detection signal is high mean curvature in DINOv2 representation space, then a generative model could be augmented with a curvature-minimization loss during training or fine-tuning — an additional term in the objective function that penalizes large angles between successive displacement vectors of the generated frames when encoded by DINOv2. Since DINOv2 is a publicly available frozen model, computing this loss term requires only forward passes through the encoder, making it computationally feasible to incorporate into existing generative training pipelines. The paper provides no evidence about whether such an attack would succeed, how much it would degrade video quality (the adversarial loss might conflict with perceptual quality objectives), or how many training iterations would be needed to close the curvature gap.

The absence of adversarial evaluation is particularly consequential given the paper's own encoder selection analysis (Figure 2B). That analysis demonstrated that different encoders produce radically different curvature gaps — DINOv2's gap of 45.46° is not a universal property of AI-generated video but a specific consequence of DINOv2's training and architecture. If a generative model were adversarially trained against DINOv2's curvature signal, it might learn to produce videos that are "straight" in DINOv2 space but would still be detectable by other encoders (or by updated versions of DINOv2). The paper does not explore whether multi-encoder ensembles could provide robustness against single-encoder adversarial attacks, nor does it characterize how transferable the curvature signal is across encoders (would a video adversarially optimized against DINOv2 also show reduced curvature in CLIP or SimCLR representations?).

The paper's suggested mitigation — "employ several detection methods in tandem, since it may be harder to game multiple metrics simultaneously without sacrificing video quality" (Section 8) — is sensible but entirely speculative. No multi-method ensemble is evaluated, no evidence is provided about the trade-off between curvature reduction and video quality, and no experiment demonstrates that gaming one metric necessarily conflicts with gaming another. The limitation is that the method's robustness to a known, clearly specified attack vector is completely uncharacterized, and a practitioner deploying ReStraV in an adversarial setting (e.g., detecting coordinated disinformation campaigns where adversaries actively seek to evade detection) has no empirical basis for estimating how quickly or effectively the method could be defeated.

Mitigation status: The attack vector is acknowledged in principle (Goodhart's law discussion) but no adversarial evaluation is conducted. The paper suggests multi-method ensembles as a mitigation strategy but provides no evidence of their effectiveness. A practitioner in an adversarial setting would need to assume that the curvature signal is vulnerable to targeted optimization by generative model developers and would need to develop their own adversarial robustness evaluation before deployment.

The Linear Relationship Between Video Length and Computational Cost Is Not Addressed, Making Long-Video Analysis Impractical Despite Its Superior Accuracy

The ablation study on video length (Appendix B.1, Figure 15a) demonstrates a clear monotonic relationship: longer analyzed durations produce higher detection accuracy. At 1 second (T ≈ 10 frames), AUROC is approximately 92%; at 2 seconds (T = 24 frames, the operating point), AUROC exceeds 96%; at 5 seconds (T = 50 frames), AUROC reaches approximately 98%. This 6-percentage-point AUROC improvement from 1 to 5 seconds is substantial, and the paper notes that "using longer videos could further enhance performance" (Section 4) — yet the method's fixed 2-second window is presented as a choice for computational efficiency rather than as a constraint that leaves accuracy on the table.

The problem is that inference time scales approximately linearly with the number of frames — each additional frame requires another DINOv2 forward pass (dominated by the 43.6 ms batch processing cost), and more frames mean more curvature and distance computations. If 5-second analysis yields ~98% AUROC but costs roughly 2.5× the inference time of the 2-second configuration (~120 ms vs. ~48 ms), a practitioner faces a sharp accuracy-efficiency trade-off that the paper does not characterize or optimize. The 2-second window is a point estimate on the accuracy-efficiency curve, not a computed optimum.

This matters because many real-world videos are significantly longer than 5 seconds — surveillance footage, lecture recordings, live streams, and user-generated content routinely run to minutes or hours. The paper's approach of analyzing a fixed 2-second window from each video means that the vast majority of temporal information in long videos is discarded. If the chosen 2-second window happens to contain an unrepresentative segment — an unusually static portion of a natural video, or an unusually coherent segment of an AI-generated video — the classification could be incorrect even though other segments would provide a clear signal. The robustness analysis on window position (Appendix B.2, Figure 16) shows that performance is relatively stable across window positions, but this analysis was conducted on 5-second videos — it does not address whether stability holds for 30-second, 5-minute, or hour-long videos where the range of possible window content is far more diverse.

The paper also does not explore whether multiple windows could be aggregated for long videos — for instance, computing curvature statistics independently for several 2-second windows and averaging or voting — which could recover some of the accuracy benefit of longer analysis without the linear cost scaling (since windows could be processed in parallel). The method's reliance on a single contiguous temporal window is an architectural constraint, not a necessity — the frozen frame-level DINOv2 embeddings are computed independently, so there is no technical barrier to analyzing multiple windows simultaneously on parallel hardware.

Mitigation status: Not addressed. The paper does not discuss the accuracy-efficiency trade-off as a function of video duration, does not propose strategies for long-video analysis (multiple windows, adaptive window selection, temporal aggregation), and does not characterize the optimal operating point on the duration-accuracy curve. The 2-second window is justified as an empirical sweet spot based on ablation studies on short (≤5s) videos, but no optimization framework is provided. A practitioner working with long videos must either accept the accuracy penalty of discarding temporal information or develop their own multi-window or adaptive analysis strategy without guidance from the paper on how to combine signals across windows or how window count affects the false positive / false negative trade-off.

7. Implications and Future Directions

How This Work Changes the Landscape

ReStraV introduces a paradigm shift in AI-generated video detection — not an incremental accuracy improvement over existing architectures, but a fundamental reframing of what the detection problem is and what kind of signal solves it. The magnitude of this shift becomes clear when contrasted with the field's trajectory prior to this work.

Before ReStraV, the dominant approach to AI video detection was to treat it as a supervised classification problem: collect training data from known generative models, train increasingly sophisticated architectures to recognize their artifacts, and hope for generalization. This produced detectors that worked reasonably well on seen generators (VideoSwin achieving 77.91% on VidProM, Table 3) but degraded sharply on unseen ones (dropping to 62.44%). The underlying logic was discriminative: learn the difference between real and fake by example. Generators improved, artifacts shifted, and detectors fell behind — an arms race where detection was permanently reactive.

ReStraV replaces this with a generative detection principle. Rather than learning to recognize what fake videos look like, it measures whether a video's temporal dynamics conform to the statistical expectations of a representation that was trained exclusively on natural data. The representation — DINOv2, frozen and never exposed to synthetic content — serves as a normative model of natural temporal structure. Detection becomes a goodness-of-fit test: does this video's trajectory through representation space respect the geometry that natural visual statistics induce? If yes, classify as real. If the trajectory is more curved, more erratic, less predictable — classify as synthetic.

This reframing has three consequences that collectively change the detection landscape:

First, it decouples detection from generator-specific training. Because the detection criterion is deviation from natural temporal smoothness rather than similarity to known artifacts, the method generalizes to generators it has never seen — not because the classifier is robustly trained, but because the signal itself (curvature in DINOv2 space) is generator-agnostic. The evidence is in the numbers: 80.05% accuracy on Sora with zero Sora videos in training (Table 3), 83.2% accuracy on Veo3 in zero-shot evaluation (Section 7D), and competitive one-to-many generalization where training on a single generator transfers to ten unseen ones (Table 5). These aren't small generalization improvements over baselines — they represent a qualitatively different generalization regime. A VideoSwin trained on four generators drops to near-chance on Sora (60.70%); ReStraV drops to 80.05%. The difference is not a better architecture — it's a better question. "Does this video's trajectory match natural temporal statistics?" is a more generalizable question than "Does this video contain artifacts I was trained to recognize?"

Second, it demonstrates that frozen image encoders can serve as temporal anomaly detectors without video training. This is a methodological discovery with implications beyond detection. DINOv2 was trained on static images using self-supervision — it has never seen a video, never learned to model motion, never been optimized for temporal coherence. Yet its representation space straightens natural video trajectories while leaving synthetic trajectories curved. This implies that temporal structure is implicitly learned during static image pretraining — the statistical regularities that make a sequence of frames "natural" are encoded in the geometry of the representation space even when the training objective is purely per-image. This finding suggests a broader principle: large-scale self-supervised pretraining on natural data may produce representations that are inherently sensitive to the types of temporal inconsistency that current generative models exhibit, making them ready-made anomaly detectors without any task-specific adaptation.

Third, it provides the detection field with a falsifiable theoretical framework. Most detection papers make unbounded performance claims: "our method detects AI video." ReStraV makes a conditional claim: "our method detects AI video as long as generative models fail to produce trajectories that match natural temporal statistics in DINOv2 representation space." This condition is concrete, measurable, and immediately suggests experiments to test its limits. It defines the terms of the detection-generation arms race in a way that artifact-based methods cannot: the target for generative model developers who want to evade detection is clear (reduce curvature in DINOv2 space), and the research program for improving detection is equally clear (find representation spaces or geometric measures that remain sensitive to synthetic content even when simple curvature is matched). The paper acknowledges Goodhart's law explicitly — "when a measure becomes a target, it ceases to be a good measure" (Section 8) — but by specifying the measure so precisely, it enables the field to plan for its eventual failure rather than being surprised by it.

The reconciliation of prior contradictions follows naturally from this reframing. The paper's encoder comparison (Section 4, Figure 2B) explains why some prior detection approaches worked well and others didn't: the signal was never about the sophistication of the detector architecture but about whether the underlying representation space exhibited differential straightening — a property that varies dramatically across encoders and is uncorrelated with absolute straightening capability (ρ = -0.13, p = 0.64). The spatio-temporal models (S3d, R(2+1)D, MViT) that were explicitly trained on video produce only small curvature gaps (5–20°) because their temporal training makes them robust to the very inconsistencies that provide the detection signal. The HVS-inspired models that straighten trajectories most aggressively produce negative gaps because they normalize away the differential signal. The self-supervised image models that perform best (DINOv2) do so not because they are the best visual models but because their representational geometry is selectively tuned to natural statistics — a property that emerges from training distribution and objective, not from architectural sophistication.

This reframing redirects research attention in specific ways. It makes the following directions more attractive:

Representation-centric detection research: rather than designing better temporal architectures, the focus shifts to understanding which pretrained representations produce the largest differential straightening, and why. The encoder comparison in Figure 2B is a starting point, not an endpoint — a systematic study of how pretraining data, objective, architecture, and scale affect selective straightening would be directly actionable for detection.
Geometric feature engineering: if curvature is the primary signal, what other trajectory-geometric properties might capture residual anomalies when curvature is matched? The paper mentions curvature kurtosis (Section 8, Discussion) as one candidate; others might include local curvature autocorrelation, distance-curvature interactions, or multi-scale trajectory analysis.
Adversarial robustness characterization: the falsifiable hypothesis implies a clear attack vector (curvature minimization during generation), and the paper provides no adversarial evaluation. Understanding how quickly and at what perceptual cost the curvature signal can be reduced is now a central question for the field.
Domain-specific validation of the differential straightening principle: the paper only evaluates on conventional natural video. Whether the principle extends to medical imaging, scientific visualization, surveillance, or artistic content is unknown and determines the method's deployability.

And makes these directions less urgent:

Incremental improvements to video detection architectures for within-distribution accuracy. The paper's generalization results suggest that the ceiling on artifact-based detection is determined not by architecture quality but by the fundamental brittleness of learning generator-specific signals. A 1–2% improvement on VidProM within-distribution accuracy through architectural innovation is less valuable than understanding and extending the generalization principle that ReStraV demonstrates.
Collecting ever-larger datasets of generated videos from current models. If the signal is generator-agnostic, training on more generators provides diminishing returns. The real challenge is characterizing and extending the signal's reach to future generators and new domains — which requires better representations and geometric measures, not more training data from the same model class.

Follow-Up Research This Work Enables

Systematic characterization of the relationship between pretraining data, objective, and differential straightening. The encoder comparison in Figure 2B is the most revealing experiment in the paper — and it barely scratches the surface. It tests 14 encoders and finds that DINOv2 produces the largest curvature gap (45.46°), but provides no analysis of why. Is it DINOv2's training data diversity? Its self-supervised objective (student-teacher distillation with centering and sharpening)? The ViT architecture's sensitivity to fine-grained spatial statistics? The scale of pretraining? A follow-up study would train a controlled family of encoders — varying dataset composition (natural images only, natural + synthetic, domain-restricted), training objective (contrastive, reconstructive, distillation-based, supervised), architecture (ViT vs. ConvNet vs. hybrid), and model scale — and measure the resulting curvature gap on a fixed benchmark of natural and AI-generated video pairs. The key measurement would be how the curvature gap changes as a function of each variable, and whether the gap can be predicted from properties of the training distribution (e.g., the distance between the training distribution and the natural video statistics of the test set). This would transform the encoder selection problem from empirical trial-and-error (as in Section 4) to principled design: if we know what properties cause differential straightening, we can actively construct or select encoders that maximize it.

Adversarial fine-tuning of generative models against the curvature signal. The paper's most important unfollowed thread is the falsifiable hypothesis it articulates: that future generative models could close the curvature gap by learning to produce temporally coherent trajectories in representation space. This hypothesis is testable today. A follow-up study would take an existing text-to-video diffusion model (e.g., an open-source model comparable to VideoCrafter2 or ModelScope) and augment its training with a curvature regularization loss: during the denoising process, an auxiliary loss term penalizes the expected curvature (mean θ_i in DINOv2 space) of the generated video sequence. The experiment would measure: (1) how much the curvature gap decreases as a function of regularization strength, (2) how much perceptual quality degrades (via standard VQA metrics and human evaluation), and (3) whether curvature-minimized videos remain detectable by other encoders (CLIP, SimCLR) or by multi-encoder ensembles. The critical result would be the curvature-quality Pareto frontier: the minimum achievable curvature gap at each level of perceptual quality. If even strong regularization cannot reduce the gap below detectable levels without severely degrading video quality, ReStraV's signal is robust in practice. If modest regularization closes the gap while maintaining quality, the detection-generation arms race has a clear path forward, and the field needs to develop geometric measures beyond simple mean curvature. This experiment is feasible with current open-source models and DINOv2's public availability — it requires no new architectures, only the addition of a differentiable loss term to existing training pipelines.

Multi-scale trajectory analysis and higher-order geometric features. The permutation feature importance analysis (Figure 11) reveals that mean curvature μ_θ dominates detection, with curvature variance σ²_θ as the second most important feature. But these are first- and second-order statistics of a single-scale trajectory (24 frames at ~87 ms intervals). The paper's own discussion (Section 8) suggests that future work should investigate "curvature kurtosis" and other higher-order statistics. A systematic follow-up would expand the geometric feature space to include: (1) multi-scale curvature — computing curvature at multiple temporal downsampling rates (e.g., every frame, every 3rd frame, every 9th frame) to capture anomalies at different temporal frequencies; (2) higher-order moments — skewness and kurtosis of curvature and distance distributions, which may capture sparse-event structure (the "mostly straight, occasionally sharp turn" pattern the paper observes in natural videos); (3) temporal dependency measures — autocorrelation of curvature sequences, or mutual information between curvature at different time lags; (4) geometry-quality interactions — the relationship between curvature spikes and frame-level perceptual quality metrics. The experiment would measure how much each additional geometric feature improves classification AUROC beyond the 8-statistic baseline, and whether certain combinations of features capture anomalies that survive curvature-matching adversarial attacks (tying into the adversarial fine-tuning experiment above). The goal is to understand whether curvature is the only useful geometric signal or merely the easiest to measure, and to build a more robust geometric feature set that would remain informative even if simple mean curvature were adversarially matched.

Domain-shift evaluation of ReStraV on non-photographic and specialized video domains. All of ReStraV's evaluations use natural videos that match DINOv2's photographic image training distribution: everyday scenes, human activities, natural landscapes. The paper acknowledges in its Broader Impacts that "biases inherited from pre-trained encoders may cause fairness issues across diverse content," but provides no empirical characterization. A critical follow-up would test ReStraV on deliberately out-of-distribution natural video domains: medical video (laparoscopic surgery, endoscopic examinations, ultrasound), scientific imaging (microscopy time-lapses, particle physics collision visualizations, astronomical survey sequences), artistic content (hand-drawn 2D animation, stop-motion, abstract visual music), synthetic natural video (computer-rendered architectural walkthroughs, physics simulations of natural phenomena that are not AI-generated in the paper's sense), and surveillance footage (fixed-camera, low-resolution, infrared). For each domain, the experiment would measure: (1) baseline curvature statistics of domain-natural videos in DINOv2 space, (2) whether AI-generated videos from the same domain exhibit elevated curvature relative to domain-natural videos, and (3) whether ReStraV's classifier (trained on general natural/AI video pairs) maintains acceptable false positive / false negative rates when applied to domain-specific video pairs. The critical measurement is the domain-specific curvature gap — if medical natural videos have naturally high curvature (because they contain rapid tool motions, specular highlights, or tissue deformations unfamiliar to DINOv2), then ReStraV would produce high false positive rates in that domain. Conversely, if an AI model specifically trained on medical data produces videos with domain-appropriate curvature, ReStraV would miss them. This experiment defines the deployability envelope of the method and identifies which domains require domain-specific encoders or geometric measures.

Temporal-window optimization and adaptive frame allocation. The ablation studies (Appendix B.1, Figure 15) demonstrate a monotonic relationship between analyzed duration and AUROC (92% at 1s → 96% at 2s → 98% at 5s), but the paper selects a fixed 2-second operating point without formal optimization. A follow-up would reframe this as a cost-accuracy optimization problem. The experiment would measure: (1) how classifier confidence (predicted probability from the MLP) evolves as frames accumulate — does confidence stabilize after the first few frames for easy cases, or does it require the full 24? (2) Whether an early-exit policy can be trained: given partial curvature and distance statistics from the first K frames, can a lightweight decision module reliably predict whether the full 24-frame analysis will flip the classification? (3) The trade-off between average frame count and accuracy under various early-exit policies, compared to the fixed 24-frame baseline. The key metric is the accuracy-efficiency Pareto frontier: for a target accuracy of 94%, what is the minimum average frame count achievable through adaptive allocation? The hypothesis is that many videos — both trivially real and trivially synthetic — reveal their classification early in the trajectory, and only borderline cases require the full temporal window. If true, adaptive allocation could maintain ReStraV's accuracy while reducing average inference cost by a factor comparable to the 4× efficiency gains reported in the LLM test-time compute scaling literature. This experiment is straightforward to implement with the existing pipeline: the per-frame DINOv2 embeddings are already computed sequentially, and curvature statistics can be updated online as each new frame arrives.

Temporal-direction-aware detection that captures reversal and manipulation attacks. The paper explicitly notes that ReStraV is invariant to playing a video backwards — curvature is a symmetric function of the displacement vectors. This limitation means that temporally manipulated real videos (reversed segments, frame reordering that maintains local smoothness) would evade detection. A follow-up study would investigate whether directional geometric measures can complement curvature to capture temporal anomalies that curvature misses. Candidate measures include: (1) the arrow of time in representation space — does the distribution of displacement vectors Δz_i have a preferred direction that reverses when the video is played backwards? This could be measured as the mean cosine similarity between Δz_i and a learned "forward time" direction; (2) causal dependency metrics — Granger causality or transfer entropy between representation dimensions, which should be asymmetric in time for physically plausible videos; (3) frame-ordering sensitivity — train a lightweight model to distinguish forward from reversed sequences of DINOv2 embeddings, and use its confidence as a directional signal. The experiment would evaluate whether combining curvature with directional measures improves detection of: (a) temporally reversed natural videos (which should now be flagged as anomalous), (b) AI-generated videos that happen to have low curvature (testing whether directional anomalies persist even when curvature is low), and (c) adversarially curvature-minimized videos from the experiment above. The goal is to build a detection signal that captures both "this video's frame-to-frame transitions are geometrically erratic" (curvature) and "this video's temporal dynamics respect physical causality" (direction), making it harder for adversaries to evade both signals simultaneously.

Practical Applications and Downstream Use Cases

Large-scale content moderation for video sharing platforms. A platform processing hundreds of millions of user uploads daily needs detection that is fast, cheap, and generalizes to novel generative models without constant retraining. ReStraV's 48 ms per video inference time, frozen encoder (no per-generator training required), and 97.06% accuracy on VidProM (Table 2) make it deployable as a first-pass screening filter. Videos flagged as AI-generated can be queued for human review or subjected to more expensive secondary analysis, while videos classified as natural with high confidence bypass further scrutiny. The critical operational metric is false positive rate on natural videos: ReStraV's MLP achieves 6.7% false positive rate (Figure 9B), meaning roughly 1 in 15 natural videos would be incorrectly flagged. For a platform with 100 million daily uploads, this translates to ~6.7 million false positives per day — high enough that human review is impractical, but manageable if the flag serves as a prioritization signal for automated secondary screening (e.g., metadata analysis, reverse image search for known synthetic content, or multi-encoder ensemble verification). The 80.05% accuracy on future generators like Sora (Table 3) means the filter remains partially effective against novel generative models without any update to the detection pipeline — a critical property when new models appear faster than detection systems can be retrained.

Forensic analysis of evidentiary video in legal and journalistic contexts. When the authenticity of a single video is at stake — a leaked political recording, footage submitted as evidence in court, a viral clip claimed to document an atrocity — the acceptable computational budget is higher, and the required false positive rate is near zero. In this setting, ReStraV serves not as an automated decision-maker but as a diagnostic tool that provides an interpretable signal to human analysts. Rather than outputting a binary classification, the system can report the video's mean curvature, curvature variance, and per-frame curvature trajectory, contextualized against the distribution of these statistics in natural and AI-generated reference sets. An analyst can visually inspect the curvature trajectory (as in Figure 8): is it smooth and low-variance like typical natural video, or does it exhibit the erratic peaks and elevated baseline characteristic of synthetic content? This interpretability — the ability to point to why the system flagged a video, in terms of a physically meaningful geometric property — is essential for legal and journalistic applications where decisions must be explainable and contestable. The Physics-IQ results (Figure 7) are particularly relevant here: on matched real/generated pairs of identical scenes, ReStraV's raw curvature comparison achieves 94.9–99.0% accuracy without any trained classifier, suggesting that in controlled forensic settings where natural reference footage may be available (e.g., comparing a disputed video against authenticated footage from the same camera or location), the curvature signal alone provides near-certain discrimination.

Quality assurance and training data filtering for generative model developers. Developers of video generation models need to evaluate the temporal coherence of their outputs during training and iteration. ReStraV's curvature metric provides a quantitative, reference-free measure of temporal consistency that can be computed without human evaluation and without ground-truth reference video. A generative model's training loop can monitor the mean curvature of generated samples over training iterations: decreasing curvature indicates that the model is learning to produce more temporally coherent outputs. The curvature gap relative to a natural video reference set provides a target for optimization — when the gap closes to within some tolerance, the model has achieved naturalistic temporal smoothness (and simultaneously defeated ReStraV as a detector, per the paper's falsifiable hypothesis). The 48 ms inference time means curvature can be computed on a subset of training batches without substantially slowing the training loop. Additionally, curvature statistics can serve as a filtering criterion for training data curation: generated videos with anomalously high curvature can be excluded from datasets used to train downstream models, preventing temporally incoherent synthetic data from degrading model performance — an application the paper's self-improvement discussion (Section 8) alludes to when it mentions "distilling the outputs of applying additional test-time compute back into the base LLM."

Edge deployment for camera-based authentication in consumer devices. The method's computational profile — frozen DINOv2 ViT-S/14 (22M parameters) plus a tiny MLP, 48 ms total inference on a consumer GPU — approaches the threshold for on-device deployment in smartphones and smart cameras. A camera could compute ReStraV's curvature signal on recently recorded video segments as a real-time authenticity check, flagging footage that may have been injected by malware rather than captured by the physical sensor. The key constraint is whether DINOv2 forward passes can run on mobile-class neural processing units (NPUs) within the 48 ms budget; the paper's evaluation on an NVIDIA RTX-2080 (a desktop GPU) does not establish mobile feasibility. However, the ViT-S/14 architecture is relatively small by modern standards, and distillation into an even smaller encoder (e.g., MobileViT) while preserving differential straightening — tested via the encoder selection methodology in Section 4 — could bring latency within the real-time video requirement. The invariance of curvature to temporal reversal (noted in Section 8) is a limitation here: an attacker who captures a real video and plays it backwards would not be detected by curvature analysis alone, so deployment would require supplementary directional sensing as discussed in the future directions above.

When to Prefer This Method

The paper positions ReStraV not against a single named alternative but against two entire categories — image-based detectors and video-based detectors — and the experimental design makes the trade-offs clear without requiring a formulaic matrix:

Prefer ReStraV when:

Generalization to unseen or future generative models is critical. The paper demonstrates 80.05% accuracy on Sora (Table 3) and 83.2% on Veo3 (Section 7D) with no training data from either model. Video-based detectors like VideoSwin drop to 60.70% on Sora and below 65% on unseen generators. If your deployment scenario involves a constantly evolving threat landscape where new generative models appear regularly, ReStraV's generator-agnostic signal provides a robustness baseline that artifact-based detectors cannot match.
Computational budget per video is severely constrained, and latency matters. At 48 ms per video end-to-end on a single consumer GPU, ReStraV is substantially faster than training and running dedicated video detection architectures. The paper doesn't provide baseline latency figures for comparison, but the difference in model complexity — a frozen 22M-parameter ViT plus a tiny MLP versus video models with hundreds of millions of parameters processing full spatiotemporal volumes — makes the efficiency advantage qualitative even without precise numbers.
Interpretability of the detection signal is required for downstream decision-making. The curvature metric is physically interpretable (mean angle between successive representational changes), can be visualized as a per-frame trajectory (Figure 8), and can be reported with reference distributions from natural and synthetic video corpora. This contrasts with black-box deepfake detectors that output a probability score without explaining what visual or temporal features drove the classification.
No per-generator training data is available, or training data access is restricted. Because DINOv2 is frozen and never exposed to AI-generated video, ReStraV requires only natural video data for training the lightweight classifier (or no training at all for raw curvature comparison, as in the Physics-IQ evaluation). In scenarios where generative models are proprietary or inaccessible (e.g., detecting outputs from a competitor's closed-source model), this is a decisive advantage over methods that require representative synthetic training data.

Prefer dedicated video detectors (VideoSwin, TimeSformer, DeMamba) when:

Within-distribution accuracy on known generators is the sole criterion, and generalization is not required. If you know exactly which generative models will appear in your deployment (e.g., internal monitoring of outputs from your organization's own generative models), a detector trained specifically on those models may outperform ReStraV's generator-agnostic signal. The paper does not provide this comparison (ReStraV outperforms VideoSwin even on seen generators in Table 3), but a detector trained with full access to the target generator's training distribution and optimized for within-distribution performance might close or reverse the gap.
The natural video domain differs substantially from DINOv2's photographic image training distribution. As discussed in Limitations, ReStraV's performance on medical, scientific, surveillance, or abstract artistic video is uncharacterized. In specialized domains where natural video statistics may already appear "anomalous" to DINOv2, a domain-specifically trained video detector that has learned the appropriate baseline temporal statistics would likely be more reliable. No empirical evidence supports this claim from the paper — it is a logical extrapolation from the method's dependence on DINOv2's representational geometry.
Directional temporal anomalies (reversed video, frame reordering) are part of the threat model. The paper acknowledges ReStraV's invariance to time reversal. If your detection target includes not just AI-generated video but also temporally manipulated real video (played backwards, segments rearranged), ReStraV's signal — which is symmetric under temporal reversal — provides no protection. A video detector trained to recognize directional temporal patterns (learned motion trajectories, causal flow, forward-biased dynamics) would be necessary, though the paper does not benchmark such detectors on manipulation tasks.