Source-linked AI summary
MotionPhys: Detecting AI-Generated Videos via Physical Consistency of Optical-Flow Trajectories
Haojin He, Hao Tan, Zichang Tan, Ajian Liu, Jun Wan
TL;DR
AI-generated videos can look visually smooth while exhibiting physically inconsistent motion, motivating detection beyond appearance artifacts. MotionPhys uses sparse optical-flow trajectories and multi-scale geometric features, achieving favorable cross-generator generalization and 78.5% overall accuracy on GenBuster++.
Problem
AI-generated videos may remain visually plausible over short frame sequences yet fail to preserve physical motion consistency throughout complete actions.
Method
MotionPhys treats sparse local optical-flow trajectories as physical evidence and models their geometric evolution across multiple temporal scales.
Results
78.5% overall accuracy on GenBuster++, with favorable cross-generator generalization across GenBuster++ and AIGVDBench.
Takeaways & Limitations
Physical trajectory features provide a lightweight and interpretable basis for distinguishing real and AI-generated videos beyond appearance artifacts.
Takeaways & Limitations
MotionPhys primarily models two-dimensional image-plane motion and does not explicitly represent depth variations, object interactions, or contact relationships.
Abstract
from arXiv · showhide
Modern AI video generation models can produce videos with high visual fidelity and seemingly smooth temporal transitions. However, visual realism does not necessarily imply physical motion consistency. Existing generative models mainly optimize distribution matching in pixel or latent spaces, without explicitly enforcing real-world constraints such as inertia, continuous forces, and trajectory geometry. Our experiments show that AI-generated videos remain visually plausible over short sequences of consecutive frames, yet fail to preserve physical motion consistency throughout a complete object action, resulting in systematic statistical discrepancies in their motion trajectories. Based on this observation, we introduce MotionPhys, a lightweight and interpretable framework that treats sparse motion trajectories as physical evidence rather than relying on appearance artifacts or generator-specific traces. By modeling the geometric evolution of trajectories across multiple temporal scales, MotionPhys reveals subtle motion inconsistencies that are difficult to capture with conventional visual cues and transforms them into a compact representation for efficient detection. Experiments on multiple datasets show that MotionPhys can effectively detect physical inconsistencies in generated videos and generalizes well across different video generators.
Introduction
MotionPhys addresses the gap between visually smooth AI-generated videos and physically consistent motion by analyzing sparse optical-flow trajectories across complete actions. It models trajectory geometry at multiple temporal scales to detect implicit motion anomalies and reports favorable cross-generator generalization.
- Appearance-based or generator-specific detectors may fail as overt synthesis artifacts become scarcer and can overfit to particular distributions or generators.
- MotionPhys represents videos with sparse local motion trajectories, retaining continuous motion evolution while discarding appearance information that is less relevant to physical consistency.
- The framework characterizes trajectory bending, directional change, and temporal smoothness using curvature, angular velocity, and jerk across multiple temporal scales.
- Visual continuity does not necessarily imply physical motion consistency, as AI-generated videos can preserve smooth adjacent-frame transitions while losing coordination across a complete action.
- MotionPhys captures implicit motion anomalies and achieves favorable cross-generator generalization on GenBuster++ and AIGVDBench.
Related Work
Related work has progressed from image and facial artifacts toward video-level spatio-temporal modeling as generative systems improve temporal continuity and scene structure.
- Existing detection methods use facial cues, image-level generative artifacts, and video-level spatio-temporal modeling to identify synthetic content.
- AI video generation has advanced from single-frame visual quality toward smooth object motion and scene-structure modeling over time.
- Recent generators improve visual quality, text alignment, and temporal continuity through approaches including diffusion transformers, flow matching, and large-scale training.
Method
MotionPhys constructs a video-level physical-consistency representation from local trajectories and uses it for lightweight, interpretable detection of motion inconsistencies.
- MotionPhys tracks local points at multiple temporal scales, characterizes their geometric and temporal evolution, and aggregates trajectory statistics into a video-level representation.
Video Input and Trajectory Extraction
The trajectory-extraction procedure samples video frames, initializes points using complementary spatial strategies, tracks them with pyramidal Lucas–Kanade optical flow, and extracts trajectories at three temporal scales.
- The method uniformly samples N = 48 frames over the entire input video before initializing tracking points.
- Shi–Tomasi corners provide textured-region observations, while regular grid points improve coverage in relatively smooth regions.
- Pyramidal Lucas–Kanade optical flow tracks each point frame by frame using inter-frame displacement to update its position.
- Trajectories are extracted over temporal scales of 8, 12, and 24 frames to capture motion evolution at different durations.
- At each scale, the sequence is divided into clips of length L and tracking is performed within each clip.
Differential Motion Representation
MotionPhys represents local motion trajectories through curvature, angular velocity, and jerk, then aggregates these descriptors across temporal scales into fixed-dimensional statistics.
- MotionPhys views each trajectory as a discrete temporal curve whose geometry and dynamics reflect the underlying motion process.
- Velocity and acceleration are computed by first- and second-order temporal differences, providing the basis for subsequent descriptors.
- Curvature measures local trajectory bending and can expose over-smoothed segments or abrupt local changes in generated videos.
- Angular velocity captures directional changes during trajectory traversal and complements curvature by identifying unstable or physically inconsistent turning behavior.
- Jerk measures high-order motion smoothness, with its magnitude capturing acceleration discontinuities and unnatural local transitions.
- Each variable-length physical sequence is summarized with statistical measures, aggregated across trajectories and temporal scales into a representation that reduces local tracking-error influence.The aggregation uses temporal scales of 8, 12, and 24 frames and produces a 504-dimensional video-level representation.
Classification Decision
MotionPhys uses explicitly constructed physical trajectory representations rather than implicitly learned RGB features, enabling lightweight classification with interpretable physical dimensions.
- MotionPhys represents each video with a 504-dimensional physical consistency vector built from explicitly constructed trajectory-level representations.
- The representation is compatible with conventional classifiers, with LightGBM adopted as the default classifier in the implementation.
- Classification results can be traced to geometric, turning-dynamics, and higher-order motion-smoothness dimensions.
Experiments
MotionPhys is evaluated on GenBuster++ and AIGVDBench, including cross-generator testing across diverse generation models and video settings.
- MotionPhys is evaluated on two public benchmarks: GenBuster++ and AIGVDBench.
- AIGVDBench contains over 440K videos generated by 31 recent video generation models.
- The evaluation covers text-to-video, image-to-video, and video-to-video generation paradigms.
- The comparison includes multimodal large language models, video forgery detectors, and video classification backbones.
- Detection performance is reported using accuracy and AUC, with numerical values expressed as percentages.
Main Results
MotionPhys achieves strong detection performance on GenBuster++ and competitive cross-generator performance on AIGVDBench, supporting the use of physical trajectory features for generated-video detection.
- 78.5% overall accuracy on GenBuster++ outperforms all compared methods.
- MotionPhys achieves 84.2% real-video accuracy and 72.8% fake-video accuracy on GenBuster++.
- Table 2 evaluates cross-generator generalization by training on Open-Sora and testing on unseen generators, reporting AUC and mean AUC across listed models.
- On AIGVDBench, MotionPhys achieves competitive performance across most subsets and generation paradigms.
Ablation Studies
Ablation and robustness studies show that MotionPhys benefits from complementary physical features, multi-scale trajectories, nonlinear classification, and perturbed trajectory sets.
- Temporal Scales: The full three-scale configuration achieves the highest ACC and AUC among temporal-scale configurations.
- Physical Features: The three physical feature groups are complementary within the hierarchical motion differential geometry representation.
- Temporal Scales: Multi-scale trajectory aggregation captures both local details and global motion trends.
- Classifier Comparison: Linear classifiers perform worse than conventional nonlinear classifiers on the identical extracted motion representation.
- Classifier Comparison: Several lightweight classifiers achieve strong performance, indicating that MotionPhys does not depend on a specific decision model.
Robustness to Trajectory Perturbations
MotionPhys remains stable when its trajectory set is randomly perturbed, indicating that video-level decisions depend mainly on collective physical-motion patterns rather than individual trajectories.
- Robustness to Trajectory Perturbations: AUC and ACC vary only slightly as trajectory perturbations increase from 0% to 50%.The evaluation covers random removal, addition, and replacement of trajectories.
- Robustness to Trajectory Perturbations: Stable performance across perturbation settings indicates reliance on collective trajectory distributions and overall physical-motion trends.
Conclusion
MotionPhys uses optical-flow trajectories and motion differential geometry to detect systematic physical-motion deviations in AI-generated videos, while remaining limited to two-dimensional image-plane patterns.
- Conclusion: MotionPhys derives physical-consistency features from local optical-flow trajectories using motion differential geometry.
- Conclusion: The framework provides a lightweight and interpretable approach for distinguishing real and AI-generated videos through physical motion regularities.
- Conclusion: MotionPhys does not explicitly model depth variations, object interactions, or contact relationships.Its primary scope is two-dimensional motion patterns on the image plane.