MotionPhys: Detecting AI-Generated Videos via Physical Consistency of Optical-Flow Trajectories

Haojin He; Hao Tan; Zichang Tan; Ajian Liu; Jun Wan

cs.CVAug 21, 2026

MotionPhys: Detecting AI-Generated Videos via Physical Consistency of Optical-Flow Trajectories

Authors: Haojin He, Hao Tan, Zichang Tan, Ajian Liu, Jun Wan

Organizations: State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China · Sangfor Technologies Inc., Shenzhen 518055, China

Abstract

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.

Figures & tables

Figure 1: The left panel shows representative trajectories from a soccer-kick sequence, while the right panel presents distributions of trajectory statistics computed from real and AI-generated videos in GenBuster++. Despite similar short-term visual appearance, the two classes exhibit systematic differences in motion coordination, directional variation, and high-order temporal stability.

Figure 2: Framework of the proposed MotionPhys. (a) Video Input initializes feature points, tracks their inter-frame motion, and connects the tracked positions into motion trajectories. (b) Differential Motion Representation characterizes the trajectories using curvature, angular velocity, and jerk to capture their geometric and temporal evolution. (c) Multi-Scale Statistical Aggregation summarizes the trajectory statistics over 8-, 12-, and 24-frame temporal scales and concatenates them into a 504-dimensional video-level representation. (d) Classification Decision performs the final prediction using a lightweight classifier.

Method	Real	Fake	Overall
Keye-VL-8B	95.70	11.80	53.75
MiniCPM-o 2.6	78.40	29.80	54.10
Qwen2.5-Omni-7B	81.40	31.40	56.40
Qwen2.5-VL-7B	92.60	27.70	60.15
InternVL3-8B	80.10	41.30	60.70
MiMo-VL-7B-RL	86.10	42.70	64.40

Table 1: Performance comparison on GenBuster++. Following the official evaluation protocol, MotionPhys is trained on GenBuster and evaluated on GenBuster++. Real and Fake denote the class-wise accuracies on real and AI-generated videos, respectively, while Overall reports the accuracy over the entire test set. All results are reported in percentages. Best in each column is bold, second best is underlined.

Method	I2V						T2V												V2V		Avg
Method	Easy Animate	LTX	Pyramid Flow	SEINE	SVD	Video Crafter	Acc Video	Animate Diff	CogVideo X1.5	Easy Animate	Hunyuan	IPOC	LTX	Open Sora	Pyramid Flow	Rep Video	Video Crafter	Wan 2.1	CogVideo X1.5	LTX	Avg
Video Classification Models
MViTv2	52.3	67.7	65.9	72.6	64.6	70.5	64.9	72.8	63.6	60.7	69.1	59.3	69.5	100.0	70.0	56.3	68.6	48.7	62.2	64.9	66.2
UniFormer	48.5	60.1	66.4	44.4	32.0	68.8	85.0	80.5	67.6	66.0	84.8	72.8	52.9	89.5	84.9	66.4	83.7	55.4	75.1	69.4	67.7
VideoSwin	60.4	70.9	65.6	90.7	85.1	77.1	62.9	73.7	65.6	63.0	60.8	63.7	70.1	100.0	63.9	60.2	76.3	55.0	57.5	66.5	69.5
VideoMAE	57.5	69.1	77.0	69.4	67.0	65.4	78.7	66.4	75.3	66.9	81.0	83.6	75.0	98.9	92.7	74.4	86.7	62.7	71.3	71.5	74.5

Table 2: Performance comparison on AIGVDBench under the official cross-generator evaluation protocol, where the detector is trained on Open-Sora and evaluated on all other unseen generators. Each entry reports AUC, Avg is the mean over all listed models.

Figure 3: Robustness of MotionPhys to trajectory-set perturbations on GenBuster++. The left side illustrates the changes in the trajectory set under random removal, random addition, and random replacement, while the right side reports AUC and ACC from the clean setting to a perturbation ratio of 50%.

No.	Curvature	Angular Velocity	Jerk	ACC	AUC
1	✓	✗	✗	74.10	81.20
2	✗	✓	✗	74.15	80.40
3	✗	✗	✓	73.60	80.01
4	✓	✓	✗	77.25	85.91
5	✓	✗	✓	75.80	84.50
6	✗	✓	✓	76.00	84.27

Table 3: Ablation of physical feature combinations on GenBuster++. ACC and AUC are reported for individual features, pairwise combinations, and the complete representation.

Setting	Scales	Frames	Dim.	ACC	AUC
Single	$S$	8	168	76.35	84.17
	$M$	12	168	76.65	85.49
	$L$	24	168	76.55	85.43
Pairwise	$S{+}M$	8, 12	336	76.60	86.16
	$S{+}L$	8, 24	336	77.20	86.70
	$M{+}L$	12, 24	336	77.50	86.88

Table 4: Ablation study of temporal-scale configurations on GenBuster++.

S

M

, and

L

denote trajectories spanning 8, 12, and 24 frames, respectively.

Classifier	AUC	ACC
Logistic regression	73.23	65.75
Linear SVM	73.16	66.20
RBF-SVM	81.55	72.20
$k$ -NN	81.97	72.95
Decision tree	82.41	73.65
Random forest	83.37	74.35

Table 5: Comparing representative classifiers on GenBuster++ with identical motion representation, reporting ACC and AUC.

Deepfake Detection

Deepfake detection identifies manipulated or AI-generated media to support trustworthy communication, identity verification, and forensic analysis.

See all 241 papers

PhysFlow: Physics-Aware Optical Flow for Motion Controllable Video Generation
Sep 8, 2026Cong Wang, Hanxin Zhu, Yonglin Tian +5Optical Flow Physical Consistency in Video Generation
State Key Laboratory of Multi-modal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China · Zhongguancun Academy, Beijing 100094, China +2
Physical Simulator In-the-Loop Video Generation
Mar 6, 2026Lin Geng Foo, Mark He Huang, Alexandros Lattas +3Physics-Guided Video Generation Physical Consistency in Video Generation
Max Planck Institute for Informatics, Saarland Informatics Campus · Saarbrücken Research Center for Visual Computing, Interaction and Artificial Intelligence · Singapore University of Technology and Design +2
The Pulse of Motion: Measuring Physical Frame Rate from Visual Dynamics
Mar 15, 2026Xiangbo Gao, Mingyang Wu, Siyuan Yang +4Physical Consistency in Video Generation Temporal Consistency in Video Generation
Texas A&M University · Visko Platform

Sep 8, 2026J/K move · Enter open · S save

PhysFlow: Physics-Aware Optical Flow for Motion Controllable Video Generation

Cong Wang, Hanxin Zhu, Yonglin Tian, Jiayi Luo+4

State Key Laboratory of Multi-modal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China · Zhongguancun Academy, Beijing 100094, China+2

Optical Flow Physical Consistency in Video Generation Controllable Video Generation

Open paper AskPDF arXiv

Video generation models have recently attracted substantial attention for their ability to generate visually compelling videos, yet ensuring physically consistent and plausible dynamics still remains a fundamental challenge, driving a growing line of research on physical realism in video generation. To address this challenge, motivated by the fact that physical regularities are primarily encoded in motion patterns, we propose PhysFlow, a novel two-stage framework for improving the physical plausibility of generated videos by decomposing video generation into motion-aware optical flow generation followed by motion-conditioned appearance synthesis. Specifically, PhysFlow consists of a physics-aware optical-flow video generator called PA-Flow and a flow-guided video generator called FlowRender. During the first stage, PA-Flow employs a physics-aware attention module to model how motion attributes and material properties influence global motion and local deformation, respectively, and generates an optical flow video as an explicit representation of motion. In the second stage, FlowRender leverages the decoupled motion representation as guidance to synthesize realistic textures and appearances, ultimately producing the final physically plausible video. To further support model training with explicit physical supervision, we construct PhysVideo, a physics-based video dataset generated with a physics engine and 3D-GS rendering, containing 10K foreground objects and 50K realistic video sequences with annotations of motion and material properties. Extensive experiments demonstrate that our proposed PhysFlow generates videos with superior physical plausibility while maintaining high visual fidelity compared with existing methods.

Abstract

Deepfake Detection

Explore similar work

PhysFlow: Physics-Aware Optical Flow for Motion Controllable Video Generation

Physical Simulator In-the-Loop Video Generation

The Pulse of Motion: Measuring Physical Frame Rate from Visual Dynamics