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

Huangsen Cao; Hongkang chu; Siyao Yu; Xin Ding; Jianfeng Dong; Yongwei Wang

cs.CVSep 22, 2026

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

Authors: Huangsen Cao, Hongkang chu, Siyao Yu, Xin Ding, Jianfeng Dong, Yongwei Wang

Organizations: Zhejiang University · University of the Chinese Academy of Sciences · Nanjing University of Information Science and Technology · Zhejiang Gongshang University

Abstract

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.

Figures & tables

Figure 1: Real videos exhibit smaller velocity responses than AI-generated videos, providing a discriminative cue for detection.

Figure 2: Overview of TRACE. The proposed framework learns generative trajectory representations, estimates adjacent-frame consistency through velocity differences, and constrains the resulting feature space around a real-video center.

Method	I2V						T2V												V2V		AVG
	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
	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
AI-Generated Image Detection Models
CNNSpot	76.71	75.41	89.59	91.79	97.72	53.92	58.54	39.90	78.08	64.15	68.82	85.43	95.72	100.0	96.37	75.95	72.78	57.16	71.42	71.01	76.02
UnivFD	76.39	80.11	87.65	97.70	95.79	71.21	61.32	77.83	87.53	68.99	60.68	86.78	89.47	99.99	94.24	79.78	87.59	54.28	81.94	73.14	80.62
NPR	77.27	88.64	95.04	93.58	98.07	56.26	42.96	47.70	78.34	44.83	57.74	81.03	96.69	100.0	93.63	69.02	72.69	45.70	87.58	84.71	75.57

Table 1: 12.87% AUC improvement over existing detectors across open-source video generators, with the best and second-best highlighted.

Method	Closed-Source Approaches											AVG
	Gen2	Gen3	Jimeng	Luma	Open Sora	Sora	Causvid 24fps	Kling	Pika	Vidu	Wan
AI-Generated Image Detection Models
CNNSpot	37.23	60.95	15.72	57.38	61.64	59.71	69.60	52.88	63.08	75.54	54.47		55.29
UnivFD	41.04	46.20	23.89	51.35	75.17	52.65	82.24	61.47	54.27	72.59	54.52	55.94
NPR	24.70	47.85	3.19	50.51	61.95	50.37	38.31	50.70	63.16	70.34	43.99	45.92
DDA	65.97	72.90	64.86	69.54	83.44	72.88	69.89	71.22	65.97	73.80	69.41	70.90

Table 2: 17.01% AUC improvement over existing methods on closed-source video generators, with the best and second-best highlighted.

Figure 3: ACC@5% FPR (%) on six datasets.

Figure 5: Robustness evaluation of TRACE in terms of AUC under different video distortions, including JPEG compression, Gaussian blur, and resizing.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Detection performance based on velocity representations under different noise levels. A relatively low noise level of 100 already yields strong performance, with further noise perturbation providing limited improvement.

Method	I2V						T2V												V2V		AVG
	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
	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
AI-Generated Image Detection Models
CNNSpot	61.65	61.43	76.35	78.65	91.62	52.00	54.42	49.38	65.93	55.78	57.33	73.57	86.85	97.48	89.25	64.30	61.25	53.47	59.20	57.42	67.37
UnivFD	61.70	67.80	73.82	92.10	87.57	54.23	52.45	64.25	73.33	54.80	52.67	72.63	76.12	97.48	82.90	65.25	67.53	50.63	67.87	61.88	68.85
NPR	62.75	76.82	87.33	83.27	92.90	51.58	51.62	51.28	70.42	51.80	56.12	70.08	90.15	97.50	84.77	62.48	57.30	50.85	74.67	70.38	69.70

Appendix

Table 4: 22.59% Improvement in ACC@5%FPR for Open-Source Video Generators.

Method	Closed-Source Approaches											AVG
	Gen2	Gen3	Jimeng	Luma	Open Sora	Sora	Causvid 24fps	Kling	Pika	Vidu	Wan
AI-Generated Image Detection Models
CNNSpot	58.72	65.32	57.08	62.08	64.32	62.42	63.70	65.42	66.54	70.56	62.90		63.55
UnivFD	57.66	61.40	57.76	59.58	66.56	59.62	69.70	64.92	60.66	67.62	60.48	62.36
NPR	58.68	65.78	57.04	62.58	65.32	63.48	59.40	66.98	66.80	66.52	62.04	63.15
DDA	62.34	67.70	64.40	65.68	74.04	66.00	63.86	65.60	63.60	65.64	65.90	65.89

Appendix

Table 5: 23.27% Improvement in ACC@5%FPR over Existing Methods on Closed-Source Video Generators.

Figure 7: Robustness evaluation of TRACE in terms of ACC@5% FPR under various video distortions, including JPEG compression, Gaussian blur, and resizing.

Method	I2V						T2V												V2V		AVG
	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
	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
AI-Generated Image Detection Models
CNNSpot	76.81	75.89	89.96	91.94	97.94	56.33	60.80	44.13	80.19	64.95	69.37	87.14	95.76	100.0	96.69	78.23	74.66	59.14	72.33	70.71	77.15
UnivFD	76.02	81.60	87.90	97.33	95.62	66.40	59.31	78.20	87.76	66.26	59.08	87.07	89.72	99.99	93.81	80.11	83.62	53.36	82.74	74.47	80.02
NPR	77.23	89.54	95.48	93.66	98.11	55.97	49.52	51.28	82.30	50.56	62.32	83.59	97.03	100.0	94.35	73.41	70.75	49.68	88.44	85.31	77.43

Appendix

Table 6: 14.04% Improvement in AP for Open-Source Video Generators.

Method	Closed-Source Approaches											AVG
	Gen2	Gen3	Jimeng	Luma	Open Sora	Sora	Causvid 24fps	Kling	Pika	Vidu	Wan
AI-Generated Image Detection Models
CNNSpot	34.59	57.87	25.27	50.38	56.53	52.12	59.84	54.12	60.46	71.40	50.41		52.09
UnivFD	33.66	43.14	27.79	41.46	64.83	42.55	71.76	55.33	46.37	66.24	45.75	48.99
NPR	30.31	52.14	28.97	48.46	57.79	49.64	35.55	55.81	61.10	64.36	44.63	48.07
DDA	54.58	67.46	56.80	61.17	78.01	64.11	58.89	62.29	57.29	63.70	62.31	62.42

Appendix

Table 7: 25.06% Improvement in AP over Existing Methods on Closed-Source Video Generators.

Figure 8: Robustness evaluation of TRACE in terms of AP under different video distortions, including JPEG compression, Gaussian blur, and resizing.

Figure 9: Saliency analysis. Grad-CAM visualizations of TRACE at the probe timestep $t{=}50$ on AI-generated videos. The top row shows the original frames, while the bottom row presents the corresponding saliency overlays.

Figure 10: Occlusion sensitivity analysis. Patch-occlusion sensitivity maps of TRACE, DDA, V-PVP, and STALL on AI-generated videos. The left column shows the original frames, while the right column presents the corresponding occlusion overlays, where $\Delta=\ell_{\mathrm{orig}}-\ell_{\mathrm{masked}}$ denotes the change in the predicted fake logit after patch masking. All methods share the same color scale.

Figure 11: Occlusion sensitivity analysis. Patch-occlusion sensitivity maps of TRACE, DDA, V-PVP, and STALL on AI-generated videos. The left column shows the original frames, while the right column presents the corresponding occlusion overlays, where $\Delta=\ell_{\mathrm{orig}}-\ell_{\mathrm{masked}}$ denotes the change in the predicted fake logit after patch masking. All methods share the same color scale.

Figure 12: Occlusion sensitivity analysis. Patch-occlusion sensitivity maps of TRACE, DDA, V-PVP, and STALL on AI-generated videos. The left column shows the original frames, while the right column presents the corresponding occlusion overlays, where $\Delta=\ell_{\mathrm{orig}}-\ell_{\mathrm{masked}}$ denotes the change in the predicted fake logit after patch masking. All methods share the same color scale.

Figure 13: t-SNE comparison of detector representations on AIGVDBench. Rows correspond to TRACE, DDA, V-PVP, and STALL, while columns represent Open-Sora (T2V), SVD (I2V), CogVideoX1.5 (V2V), and Jimeng (closed-source). Each panel visualizes the representations of 2,000 real test videos (blue) and 2,000 generated test videos (colored) from the corresponding source. TRACE consistently achieves clearer real–fake separation across different generator types, whereas prior methods exhibit better separation mainly on sources aligned with their training distributions and substantial overlap on unseen sources.

Explore similar work

Mind the Rift: Cross-Scale Coupling Mismatch for AI-Generated Video Detection
Sep 1, 2026Siyu Li, Jin Yang, Weiheng LiangDigital Image Forensics Deepfake Detection
School of Cyber Science and Engineering Sichuan University Chengdu, Sichuan, China · School of Information Science and Technology Xizang University Lhasa, Xizang, China
Retrieval-Driven Training-Free AI-Generated Video Attribution
Jul 31, 2026Renxi Cheng, Chaolei Han, Jie Gui +1Deepfake Detection
School of Cyber Science and Engineering, Southeast University, Nanjing 210096, China · Purple Mountain Laboratories, Nanjing 210000, China · Engineering Research Center of Blockchain Application, Supervision And Management (Southeast University), Ministry of Education, China +2
ATSS: Detecting AI-Generated Videos via Anomalous Temporal Self-Similarity
Apr 5, 2026Hang Wang, Chao Shen, Lei Zhang +1Deepfake Detection Multimodal Anomaly Detection
Xi’an Jiaotong University, Xi’an, China · The Hong Kong Polytechnic University, Hong Kong, China · School of Engineering and Technology, University of Washington, Tacoma, WA 98402, USA

Sep 1, 2026J/K move · Enter open · S save

Mind the Rift: Cross-Scale Coupling Mismatch for AI-Generated Video Detection

Siyu Li, Jin Yang, Weiheng Liang

School of Cyber Science and Engineering Sichuan University Chengdu, Sichuan, China · School of Information Science and Technology Xizang University Lhasa, Xizang, China

Digital Image Forensics Deepfake Detection

Open paper AskPDF arXiv

As AI video generators achieve cinematic realism, reliable detection becomes essential for safeguarding digital trust. We identify cross-scale coupling mismatch as a new forensic signal, where scale refers to the level of abstraction (semantic dynamics vs. pixel-level residuals): in natural videos, macro-level temporal dynamics and micro-level residual patterns are intrinsically coupled by the unified imaging physics pipeline, whereas AI generators, whose training objectives do not explicitly preserve this joint distribution, systematically violate this coupling. Detecting such mismatch is challenging because it requires independently extracting information at both scales while simultaneously quantifying their cross-scale relationship. We propose RIFT (Representation Inconsistency Forensics on Trajectories), an orthogonal forensic framework that addresses this through three interlocking components: a macro stream that builds a dynamic baseline of expected temporal evolution via differential geometry and persistent homology on learned manifold trajectories, a micro stream that acts as a sensitive forensic probe via steganalytic filtering and temporal modeling, and a coupling divergence module that measures the conditional dependency between the two streams. Gram-Schmidt orthogonality guarantees the information-theoretic validity of this measurement. Experiments on two benchmarks (VidProM, 120K videos, 7 generators; GenVidBench, 68K videos, 4 generators) demonstrate that RIFT achieves 99.33% and 99.72% F1-score respectively, with 97.87% unseen-generator detection rate in leave-one-out evaluation, while exhibiting encoder agnosticism: scaling from ViT-S/14 (22M) to ViT-L/14 (300M) changes F1 by less than 0.1%, and switching to a different encoder family (DINOv1) reduces F1 by only 0.73 pp. Code is available at https://github.com/Litsay/RIFT

Abstract

Explore similar work

Mind the Rift: Cross-Scale Coupling Mismatch for AI-Generated Video Detection

Retrieval-Driven Training-Free AI-Generated Video Attribution

ATSS: Detecting AI-Generated Videos via Anomalous Temporal Self-Similarity