Did You Steal My Shot? Pioneering Camera Motion Plagiarism Detection in Generative Videos

Chengguo Zhang; Ping Ping

cs.CVSep 8, 2026

Did You Steal My Shot? Pioneering Camera Motion Plagiarism Detection in Generative Videos

Authors: Chengguo Zhang, Ping Ping

Organizations: Hohai University Nanjing, China

Abstract

Camera motion often reflects directorial intent and requires professional equipment, making it a high value form of intellectual property. However, generative video models can imitate such high value camera motions with simple prompts, while existing similarity detection methods mainly operate on visual content and fail to capture deeper motion similarity. This is mainly because their training data entangles camera motion with visual content. Moreover, traditional optical flow is insufficient to represent complex camera motions. We therefore build the first benchmark for camera motion analysis, including a motion dataset with 11 motion styles and evaluation protocols. Furthermore, we propose a motion representation that augments optical flow with vorticity cues from fluid dynamics, thereby better capturing motions. Experiments show that our detector achieves a 3.02* improvement in plagiarism detection over the strongest baseline and remains effective on generative videos. We believe our work extends copyright protection beyond static content to dynamic camera motion.

Figures & tables

Figure 1. Illustration of the Camera Motion Plagiarism. High-value camera motion tends to require specialized equipment and coordination (Top), yet generative videos can easily replicate it with a simple prompt (Bottom), posing a new threat.

Figure 2. Overview of the proposed dataset generation pipeline. (a) and (b): Scenes are constructed using simple geometry and camera motions. (c) 2D coordinates and depth of each pixel at frame $t$ are obtained via ray casting. (d) and (e): Pixel positions at frame $t+1$ are computed via reprojection and occlusion check, and optical flow is derived from coordinate differences.

Category	Motion Type	DoF
Static	Static	None
Translation	Truck, Pedestal, Dolly in	$t_{x}$ , $t_{y}$ , $t_{z}$
Rotation	Pan Tilt, Roll	$r_{x}+r_{y}$ , $r_{z}$
Composite	Bullet Time, Crane, Whip Pan, Dolly Zoom, FPV Drone	$t_{x}+r_{y}$ , $t_{y}+t_{z}$ , $r_{y}+r_{z}$ , $t_{z}$ +zoom, full 6 DoF

Table 1. The 11 simulated camera motion types in the CineFlow dataset.

Figure 3. Optical flow and vorticity across 11 camera motion styles in CineFlow. First row: optical flow visualizations for one representative clip per motion style, rendered with the standard color coding. Second row: the corresponding vorticity maps computed from the same flow fields.

Figure 4. Optical flow magnitude (left) and vorticity map (right) under a roll motion.

Figure 5. Magnitude (left) and vorticity (right) distributions under a roll motion.

Figure 6. Architecture of the motion plagiarism detection network. The input flow $(u,v)$ is first augmented with vorticity $\zeta$ . Next, a 3D convolution extracts tubelet embeddings, and a learnable CLS with positional embeddings is added. Then, the embedding sequence is fed into a ViT backbone with GC adapters. Finally, the global embedding is projected.

Figure 7. Camera motion separability on CineFlow. Left: t-SNE visualization of embeddings extracted from test clips. Right: style similarity heatmap based on cosine similarity.

Figure 8. Examples and similarity analysis on camera motion models. Left: frame sequences for three motion styles, with the corresponding camera trajectories. Right: the similarity heatmap (top) and the t-SNE visualization (bottom) computed from the generated videos. Additional examples are provided in the Appendix.

Model	VideoMAE2	ViCLIP	DINO v3	VideoMamba	Ours
MotionCtrl	0.1533	0.0715	0.1261	0.2024	0.6470
CameraCtrl	0.1692	0.0598	0.0951	0.1748	0.5755
DualCamCtrl	0.1408	0.0722	0.1137	0.1515	0.5913
ReCamMaster	0.1524	0.0783	0.1496	0.1773	0.6201

Table 2. NMI

\uparrow

score of clustering on videos generated by camera motion models.

Model	VideoMAE2	ViCLIP	DINO v3	VideoMamba	Ours
MotionCtrl	-0.0762	-0.1094	-0.0868	-0.0316	0.3739
CameraCtrl	-0.0847	-0.1251	-0.1033	-0.0539	0.3486
DualCamCtrl	-0.0735	-0.1327	-0.0979	-0.0101	0.3844
ReCamMaster	-0.0880	-0.1288	-0.0942	-0.0672	0.3717

Table 3. ARI

\uparrow

score of clustering on videos generated by camera motion models.

Figure 9. Performance of plagiarism detection on generative videos. Left: cosine similarity between videos. Right: Precision@5 for retrieving videos with the same motion label.

Figure 10. Motion plagiarism detection on commercial generative videos. Given a reference camera motion (left), Veo 3.1 and Jimeng are prompted to imitate the motion while generating a new scene (middle). CineGuard outputs a motion plagiarism score based on cosine similarity (right). Additional examples are provided in the Appendix.

Figure 11. Robustness under common post processing attacks. The figure includes Random Crop, Frame Drop, Gaussian Noise, Gaussian Blur, H.264/AVC, and H.265/HEVC, for four representative camera motion styles.

Dataset	VideoMAE2	ViCLIP	DINO v3	VideoMamba	Ours
w/o CineFlow	0.2594	0.2228	0.2367	0.2311	0.3621
w/ CineFlow	0.5427	0.4783	0.5032	0.5265	0.7828

Table 4. Cosine similarity

\uparrow

under ablation on CineFlow.

Dataset	VideoMAE2	ViCLIP	DINO v3	VideoMamba	Ours
w/o vorticity	0.5427	0.4783	0.5032	0.5265	0.5371
w/ vorticity	0.6383	0.5541	0.5806	0.6074	0.7828

Table 5. Cosine similarity

\uparrow

under ablation on vorticity.

Camera Motion Understanding

See all 19 papers

TRACE: Trajectory Representation and Consistency Estimation for AI-Generated Video Detection
Sep 22, 2026Huangsen Cao, Hongkang chu, Siyao Yu +3Deepfake Detection Video Forensics
Zhejiang University · University of the Chinese Academy of Sciences · Nanjing University of Information Science and Technology +1
Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods
Mar 16, 2026Omer Ben Hayun, Roy Betser, Meir Yossef Levi +2
Technion – Israel Institute of Technology, Haifa, Israel
Detect Early, Escalate Rarely: Anytime Detection of AI-Generated Video from the Compressed Bitstream
Jul 21, 2026Mert Onur Cakiroglu, Mehmet Dalkilic, Hasan KurbanDeepfake Detection Sequential Hypothesis Testing
Luddy School of Informatics, Computing, and Engineering, Indiana University Bloomington, Bloomington, IN, USA · College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar

Sep 22, 2026J/K move · Enter open · S save

TRACE: Trajectory Representation and Consistency Estimation for AI-Generated Video Detection

Huangsen Cao, Hongkang chu, Siyao Yu, Xin Ding+2

Zhejiang University · University of the Chinese Academy of Sciences · Nanjing University of Information Science and Technology+1

Deepfake Detection Video Forensics

Open paper AskPDF arXiv

Recent advances in generative video models have enabled the synthesis of visually realistic content, posing significant challenges to synthetic video detection. Existing detectors often rely on appearance artifacts, semantic inconsistencies, and temporal patterns that may be generator-specific, limitating generalization to unseen synthesis models. We investigate whether responses to a pretrained generative model provide more transferable forensic cues. Our key observation is that real and AI-generated videos exhibit distinct velocity responses under a pretrained Flow Matching video model. This distinction persists when different pretrained video-generation backbones are used as probes, suggesting that velocity responses offer transferable forensic signals beyond visual artificts. Motivated by this observation, we propose TRACE (Trajectory Representation and Consistency Estimation), a generation-process-aware framework for AI-generated video detection. TRACE leverages a pretrained video DiT as a velocity-field probe to extract representations at multiple flow time points, and models cross-frame consistency through velocity differences between adjacent frames. We further introduce a Real-Centered Trajectory Optimization objective that encourages generator-invariant representation learning. Extensive experiments on AIGVDBench demonstrate that TRACE generalizes effectively across diverse generators, substantially outperforming prior state-of-the-art methods on unseen open- and closed-source video generation models.

Abstract

Camera Motion Understanding

Explore similar work

TRACE: Trajectory Representation and Consistency Estimation for AI-Generated Video Detection

Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods

Detect Early, Escalate Rarely: Anytime Detection of AI-Generated Video from the Compressed Bitstream