Pith. sign in

REVIEW 5 major objections 5 minor 3 cited by

GenWorld: Towards Detecting AI-generated Real-world Simulation Videos

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read High-quality AI-generated videos still break 3D consistency, and a detector that checks it catches them where pixel-based detectors fail.

desk verdict A genuinely useful benchmark and a striking Cosmos result, but the 3D-consistency explanation isn't yet backed by controls or ablations. read the letter →

arxiv 2506.10975 v1 pith:GTLWLAQG submitted 2025-06-12 cs.CV

Weiliang Chen , Wenzhao Zheng , Yu Zheng , Lei Chen , Jie Zhou , Jiwen Lu , Yueqi Duan This is my paper
classification cs.CV
keywords AI-generatedvideodetectionworldmodelsmulti-viewconsistencyDust3Rforensicscross-promptevaluationGendatasetSpannDetector
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that the weak point of high-quality AI-generated video is not pixel-level texture but three-dimensional consistency across frames. It introduces GenWorld, a 100,000-video benchmark built around real-world scenes such as driving, navigation, and manipulation, where fake videos come from ten generators spanning text-to-video, image-to-video, and video-to-video prompts. It then claims that existing detectors, which look for spatial or temporal artifacts, collapse on videos generated by world models like Cosmos, scoring around 16% to 30% accuracy, while a detector that scores multi-view geometric consistency reaches 72.44% on the same videos and the best average accuracy overall. A sympathetic reader would take this as evidence that physical plausibility is a usable forensic cue, not just an aesthetic property.

What carries the argument

The central object is the projection residual $R = \|I_1 - I_{2 \to 1}\|$, computed by projecting a frame $I_2$ onto the coordinate frame of frame $I_1$ through the point maps and camera parameters estimated by Dust3R. Real videos keep this residual small and structured because a consistent 3D scene explains both views; generated videos produce large irregular residuals because the frames do not share a coherent underlying geometry. SpannDetector feeds frame pairs to the frozen Dust3R encoder, reads a spatial memory module inspired by Spann3R to propagate information across time, and uses a learned authenticity scorer on the decoded features, with the final video score being the global average over frames.

What would settle it

Record real videos and the same videos after heavy downscaling and re-compression to match the resolution and bitrate of Cosmos-generated clips, then compute the Dust3R residual for both; if the residual statistics become indistinguishable, the reported gap reflects video quality rather than multi-view inconsistency. Alternatively, train SpannDetector on real videos whose frames are artificially warped to break multi-view consistency; if the detector then fails to transfer to genuine AI-generated videos, the proposed mechanism is not what is being learned.

Watch

Extended reading notes

Core claim

The paper's central claim is that AI-generated videos, even when visually hyper-realistic, fail to satisfy the multi-view consistency that real videos satisfy because they lack a single coherent 3D structure underlying the frames. Using the Dust3R stereo reconstruction model, the paper shows that the residual between a frame and the projection of a neighboring frame onto it is regular and structured for real videos but irregular and noisy for generated videos. On this basis, SpannDetector, built by attaching a trainable authenticity scorer and a temporal memory to the frozen Dust3R encoder-decoder, outperforms state-of-the-art detectors: on Cosmos-generated videos it reaches 72.44% accuracy versus 16.17% for DeMamba and 30.38% for VideoMAE, and it achieves the highest average accuracy of 89.06% in the train-test evaluation. The paper also shows that SpannDetector generalizes across prompt types better than DeMamba, which overfits to a single generation prompt.

Load-bearing premise

The paper assumes that the residual between a frame and the projection of a neighboring frame, as computed by Dust3R, isolates 3D inconsistency rather than being driven by differences in resolution, frame rate, compression, or generator-specific textures, and no control experiment in the paper rules out those confounds.

Editorial extensions

If this is right

  • Detectors that rely on multi-view consistency should remain effective as generation quality improves, because the physical-inconsistency cue is not a texture artifact that generators can easily erase.
  • Benchmarks for AI-generated video detection should include real-world simulation videos; otherwise models trained on cartoon-heavy data will be tested on the wrong distribution.
  • World-model outputs (video-to-video generators like Cosmos) are a harder detection target than text-to-video or image-to-video outputs, so they should be a standard stress test for new detectors.
  • Cross-prompt evaluation (training on text-to-video, testing on image-to-video and video-to-video) shows that multi-view consistency transfers better than pixel-space artifacts, making generalization a realistic goal.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the 3D-consistency cue is truly what drives detection, then any generator that learns to produce multi-view-consistent videos, for example by incorporating explicit 3D supervision, would close the gap and force detectors to find other physical cues such as dynamics or lighting.
  • The same residual signal could be measured on single videos at larger temporal distances, not just adjacent frames, potentially making the cue more sensitive to long-range inconsistencies.
  • The uncontrolled confounds of resolution, frame rate, and compression mean the reported gap in residual statistics could be smaller than it appears; a controlled experiment resizing and compressing real videos to match generator settings would test whether 3D consistency rather than low-level video quality is the operative signal.
  • The GenWorld evaluation protocol, which tests on completely unseen generators, is a stronger generalization test than the usual within-generator split; adopting it as a community standard would push methods toward learning physical priors rather than memorizing generator fingerprints.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The manuscript introduces GenWorld, a 100,296-video dataset for AI-generated video detection that focuses on real-world simulation, assembled from real videos in driving, navigation, manipulation, and human activity, and from generated videos produced by ten models spanning text-to-video, image-to-video, and video-to-video generation. The authors evaluate existing detectors on this dataset and report that they perform poorly on videos from the world model Cosmos. They then propose SpannDetector, which combines a frozen DUSt3R stereo reconstruction encoder/decoder with a trainable memory module and an authenticity scorer, and report that it improves detection, especially on Cosmos videos (72.44% versus 16.17% for DeMamba in Table 3). The paper argues that the detector leverages multi-view (3D) consistency as a physical-plausibility prior.

Significance. If the empirical claims are sustained, GenWorld is a valuable benchmark because it targets deceptive real-world simulation videos, includes cross-prompt diversity, and uses held-out generator splits that reduce circularity in evaluation. The 3D-consistency prior for AI-generated video detection is a timely and plausible direction. The dataset construction, the use of multiple state-of-the-art generators, and the user study are concrete contributions. However, the reported results are not yet supported: the evaluation metric is ambiguously defined, the residual analysis in Section 4.2 does not control for source-specific low-level differences, the final detector is not shown to use the residual cue, and several tables contain internally inconsistent aggregate numbers. These issues are load-bearing for the paper's central claims.

major comments (5)
  1. [Section 5.1 and Table 3] The definition of "Accuracy" is ambiguous and appears to be recall on generated videos only. The text says that when calculating Accuracy, the test data from the generative model itself are used, which excludes the real-video test set. Indeed, the Table 3 values for VideoMAE on HotShot (97.56) match the Recall value in Table 4 (0.9756). This means the headline numbers are not balanced classification accuracy, and comparisons with prior work that report true accuracy on real plus fake videos are not meaningful without clarification. Please report standard accuracy on the combined real and generated test sets, or clearly label the metric as recall and provide balanced accuracy as well.
  2. [Section 4.2 and Table 1] The motivating observation that residual R is larger for generated videos does not control for source-specific low-level differences. Table 1 shows that real videos range from 224 to 1600 pixels and 10 to 30 fps, while all generated videos are 8 fps and have fixed resolutions such as 256x256 or 1280x2048. A pretrained stereo network like DUSt3R, trained on natural images, can produce larger residuals on lower-resolution, re-encoded, or domain-shifted inputs regardless of 3D consistency. Figure 3(b) is only qualitative. Please provide aggregate residual statistics with confidence intervals, and matched controls such as downsampling real videos to the resolution and frame rate of generated videos, or adding equivalent compression, before attributing the difference to physical inconsistency.
  3. [Section 4.3, Eqs. (2)-(6)] SpannDetector never explicitly computes the residual R that Section 4.2 identifies as the discriminative cue. The scorer operates on DUSt3R decoder features f_d, and the memory mechanism uses attention over encoder features. Nothing in Eqs. (2)-(6) ties the final score s_t to the projection residual R. Consequently, the reported improvement on Cosmos videos could be driven by any low-level cue present in the frozen DUSt3R features, such as texture, compression, or resolution artifacts, rather than by 3D inconsistency. Please add an ablation that isolates the 3D-consistency signal, for example by replacing DUSt3R with a generic frozen video encoder, by feeding R explicitly to the scorer, or by correlating s_t with R on a per-video basis.
  4. [Table 4 and Table 5] The Avg. columns in Tables 4 and 5 are internally inconsistent. In Table 4, the F3Net Recall row shows per-generator values 0.6325, 0.5961, 0.5822, 0.3091, 0.5574, whose mean is approximately 0.535, yet the Avg. column reports 0.3657; the same Avg. value 0.3657 appears for NPR and STIL Recall rows. Similarly, the Ours F1 Avg. is 0.9322, while the mean of the five per-generator F1 values is approximately 0.850. In Table 5, the Ours T2V row reports 99.05, 51.66, and 52.37 with an Avg. of 86.09, though the row mean is 67.69. These discrepancies make it impossible to verify the aggregate claims in Section 5.3. Please recompute and correct all aggregate columns, and state precisely what quantity the Avg. column represents.
  5. [Section 5.1 and Section 5.3] No error bars, number of runs, or statistical significance tests are reported for any of the accuracy, F1, or AP results. The central claim that SpannDetector outperforms baselines rests on point estimates from a single split. Given the large reported margins, such as 72.44% versus 16.17% on Cosmos, the qualitative conclusion may survive repeated runs, but the manuscript needs at least multiple seeds with standard deviations, and ideally a paired test across methods, before the comparison is trustworthy.
minor comments (5)
  1. [Section 5.5] The user study reports means on a 7-point Likert scale without standard deviations, inter-rater agreement, or a description of how the 120 participants were assigned to the 100 videos per dataset. Including this information would strengthen the claim that the GenWorld generated videos are the most deceptive.
  2. [Table 1] The Resolution column uses inconsistent notation, such as "224-340" for Kinetics and "960-540" for DL3DV-10K, and it is unclear whether these are width-height pairs or ranges. Please standardize the format and specify the units.
  3. [Table 3] The VideoMAE row is difficult to read because values are concatenated without spaces, for example "97.5630.3893.1581.02 99.59". Please format all rows with consistent spacing.
  4. [Section 4.1, Eq. (1)] The superscript notation X^{1,1}, X^{2,1} is unnecessarily confusing; using subscripts such as X_{1,1} and X_{2,1} throughout would improve readability.
  5. [Section 2] VideoMAE is a self-supervised video representation model rather than a dedicated AI-generated video detector. Please describe how it is adapted for detection, since this affects the interpretation of its performance in Tables 3 and 4.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark results are empirical, evaluated on held-out generators, and the multi-view consistency story is a motivating observation rather than a fitted input repackaged as a prediction.

full rationale

The paper's central claims are empirical benchmark results. SpannDetector is trained on the training split (five generators plus real data) and tested on held-out test sets from five other generators (Section 5.1), so the reported Cosmos accuracy (72.44% vs. DeMamba's 16.17% in Table 3) is not a fitted parameter renamed as a prediction. The multi-view residual analysis (Section 4.2) is used as motivation for the architectural choice of a frozen Dust3R encoder/decoder with a trained memory and authenticity scorer (Eqs. 2-6), but the final scores are learned and evaluated on unseen generators; no equation defines the test accuracy in terms of the motivating residual R. Self-citations (OccWorld [47], OWL-1 [23], DreamCinema [13]) are contextual related-work references and are not load-bearing for the detector's design or the benchmark conclusions. The main scientific caveat, that the Cosmos improvement may be driven by resolution, framerate, compression, or domain shift rather than by 3D inconsistency, is a validity and attribution threat rather than a circularity. Under the specified criteria, no circular step can be exhibited with a direct equation-to-equation or fitted-input-to-prediction reduction.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The claims rest on the dataset's design premise, on Dust3R's reliability as a geometric prior, and on the assumption that observed residual differences reflect physical inconsistency rather than distribution shift. The learned scorer and memory introduce fitted parameters whose training is not fully specified.

free parameters (1)
  • SpannDetector scorer and memory weights = not reported
    The authenticity scorer and memory update are trained on the GenWorld training split; without training details or model weights, the fitted nature of the detector is opaque, though evaluation on held-out generators mitigates overfitting concerns.
assumptions (5)
  • domain assumption Videos that closely mimic real-world scenarios are the most impactful and should be the focus of forgery detection.
    Section 3.1 motivates GenWorld with this premise, which determines the dataset's scope and the evaluation's relevance.
  • domain assumption The Dust3R point-map residual R = |I1 - I2,1| reliably measures physical multi-view consistency, and generated videos have larger, noisier residuals.
    Section 4.2 assumes this without quantitative statistics or controls for resolution, frame rate, and compression differences between real and generated videos.
  • domain assumption Frozen Dust3R encoder/decoder features, augmented by memory, retain the consistency cue and can be scored for authenticity.
    Section 4.3 builds SpannDetector on this assumption; no ablation isolates the contribution of each component.
  • standard math Dust3R provides valid point maps and camera intrinsics for arbitrary frame pairs.
    Section 4.1 relies on the pretrained Dust3R as ground truth for multi-view geometry.
  • domain assumption Held-out generator splits measure generalization rather than memorization.
    Section 5.1 defines train/test splits by generator; the assumption is that no content overlap or source leakage exists between real and generated videos.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GenWorld: Towards Detecting AI-generated Real-world Simulation Videos." pith.science (2026). https://pith.science/paper/GTLWLAQG

@misc{pith2026250610975,
  author       = {Pith},
  title        = {Pith review of: GenWorld: Towards Detecting AI-generated Real-world Simulation Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GTLWLAQG}},
  note         = {Machine review of arXiv:2506.10975}
}
read the original abstract

The flourishing of video generation technologies has endangered the credibility of real-world information and intensified the demand for AI-generated video detectors. Despite some progress, the lack of high-quality real-world datasets hinders the development of trustworthy detectors. In this paper, we propose GenWorld, a large-scale, high-quality, and real-world simulation dataset for AI-generated video detection. GenWorld features the following characteristics: (1) Real-world Simulation: GenWorld focuses on videos that replicate real-world scenarios, which have a significant impact due to their realism and potential influence; (2) High Quality: GenWorld employs multiple state-of-the-art video generation models to provide realistic and high-quality forged videos; (3) Cross-prompt Diversity: GenWorld includes videos generated from diverse generators and various prompt modalities (e.g., text, image, video), offering the potential to learn more generalizable forensic features. We analyze existing methods and find they fail to detect high-quality videos generated by world models (i.e., Cosmos), revealing potential drawbacks of ignoring real-world clues. To address this, we propose a simple yet effective model, SpannDetector, to leverage multi-view consistency as a strong criterion for real-world AI-generated video detection. Experiments show that our method achieves superior results, highlighting a promising direction for explainable AI-generated video detection based on physical plausibility. We believe that GenWorld will advance the field of AI-generated video detection. Project Page: https://chen-wl20.github.io/GenWorld

Figures

Figures reproduced from arXiv: 2506.10975 by the authors.

Figure 1
Figure 1. Most existing AI-generated video datasets consist of cartoon videos even as “real” data, lacking a clear definition of authenticity. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Video Generation Pipeline of GenWorld. 3.2. Collection and Organization of GenWorld Given the above analysis, we aim to build a high-quality, real-world AI-generated video detection dataset. Firstly, for real-world data, we analyze real-world scenarios and the current video generation world models [3, 23], select￾ing four key scenarios that capture a broad spectrum of hu￾man life: driving, indoor and outdoor navigat… view at source ↗
Figure 3
Figure 3. Pipeline and motivation of Our SpannDetector. SpannDetector is designed based on an in-depth analysis of multi-view consistency in real and AI-generated videos. It integrates a stereo reconstruction model with a temporal memory module to enhance efficiency in consistency detection. An authenticity scorer evaluates the stereo features, and the final video authenticity is determined by averaging these scores across th… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: More Visualization Results of our dataset. more comprehensive and accurate evaluation. Baseline. We select state-of-the-art methods in AI￾generated content detection, including both image [32, 35] and video [11, 20, 36, 44] detectors, and evaluate their per￾formance. A…
Figure 5
Figure 5. Figure 5: Temporal Visualization Results of our dataset [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SphereVideo: Prototype-anchored Hyperspherical Boundary for Continual AI-generated Video Detection

    cs.CV 2026-08 conditional novelty 6.0 of 10

    A hyperspherical prototype boundary with temporal-coherence losses improves continual AI-generated video detection by about 3 to 4 percentage points over prior methods.

  2. SafeGuard: A Multi-Agent Perception-Reasoning Framework for Social-Risk AI-Generated Video Detection

    cs.CV 2026-07 conditional novelty 6.0 of 10

    SafeGuard bridges low-level forensic perception and high-level semantic reasoning via multi-agent collaboration, lifting accuracy +18.7% on a new social-risk AI-video benchmark.

  3. Measuring 3D Spatial Geometric Consistency in Dynamic Video Generation

    cs.CV 2026-03 conditional novelty 6.0 of 10

    SGC quantifies 3D geometric consistency of generated videos by measuring divergence among local camera poses estimated only on static background sub-regions.

Reference graph

Works this paper leans on

48 extracted references · 20 canonical work pages · cited by 3 Pith papers

  1. [1]

    3, 4, 5, 7, 8

    Hotshot, 2023. 3, 4, 5, 7, 8

    work page 2023

  2. [2]

    2, 3, 4, 5

    Zeroscope-v2-xl, 2024. 2, 3, 4, 5

    work page 2024

  3. [3]

    Cosmos world foun- dation model platform for physical ai.arXiv preprint arXiv:2501.03575, 2025

    Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foun- dation model platform for physical ai.arXiv preprint arXiv:2501.03575, 2025. 1, 2, 3, 4, 5, 7, 8, 9

    arXiv 2025

  4. [4]

    Ai-generated video detection via spatial-temporal anomaly learning

    Jianfa Bai, Man Lin, Gang Cao, and Zijie Lou. Ai-generated video detection via spatial-temporal anomaly learning. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 460–470. Springer, 2024. 2, 6

    work page 2024

  5. [5]

    Navigation world models.arXiv preprint arXiv:2412.03572, 2024

    Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models.arXiv preprint arXiv:2412.03572, 2024. 2, 3

    arXiv 2024

  6. [6]

    Identify- ing and mitigating the security risks of generative ai.Founda- tions and Trends® in Privacy and Security, 6(1):1–52, 2023

    Clark Barrett, Brad Boyd, Elie Bursztein, Nicholas Car- lini, Brad Chen, Jihye Choi, Amrita Roy Chowdhury, Mihai Christodorescu, Anupam Datta, Soheil Feizi, et al. Identify- ing and mitigating the security risks of generative ai.Founda- tions and Trends® in Privacy and Security, 6(1):1–52, 2023. 2

    work page 2023

  7. [7]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023. 2

    arXiv 2023

  8. [8]

    Rt-1: Robotics transformer for real-world control at scale

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakr- ishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3, 4

    arXiv 2022

Show all 48 references
  1. [9]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luh- man, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators

  2. [10]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and patter...

    2020

  3. [11]

    Demamba: Ai-generated video detection on million-scale genvideo benchmark.arXiv preprint arXiv:2405.19707, 2024

    Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. Demamba: Ai-generated video detection on million-scale genvideo benchmark.arXiv preprint arXiv:2405.19707, 2024. 2, 3, 4, 5, 6, 7, 8

    2024 arXiv

  4. [12]

    Videocrafter2: Overcoming data limitations for high-quality video diffusion models.arXiv preprint arXiv:2401.09047, 2024

    Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models.arXiv preprint arXiv:2401.09047, 2024. 2, 3, 4, 5, 7, 8

    2024 arXiv

  5. [13]

    Dreamcinema: Cinematic transfer with free camera and 3d character.arXiv preprint arXiv:2408.12601, 2024

    Weiliang Chen, Fangfu Liu, Diankun Wu, Haowen Sun, Haixu Song, and Yueqi Duan. Dreamcinema: Cinematic transfer with free camera and 3d character.arXiv preprint arXiv:2408.12601, 2024. 2

    2024 arXiv

  6. [14]

    Seine: Short-to-long video diffu- sion model for generative transition and prediction

    Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffu- sion model for generative transition and prediction. InThe Twelfth International Conference on Learning Representa- tio...

    2023

  7. [15]

    The deepfake detection challenge (dfdc) dataset.arXiv preprint arXiv:2006.07397, 2020

    Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset.arXiv preprint arXiv:2006.07397, 2020. 2

    2006 arXiv

  8. [16]

    Privacy and security concerns in generative ai: a compre- hensive survey.IEEE Access, 2024

    Abenezer Golda, Kidus Mekonen, Amit Pandey, Anushka Singh, Vikas Hassija, Vinay Chamola, and Biplab Sikdar. Privacy and security concerns in generative ai: a compre- hensive survey.IEEE Access, 2024. 2

    2024

  9. [17]

    Generative adversarial networks.Commu- nications of the ACM, 63(11):139–144, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Commu- nications of the ACM, 63(11):139–144, 2020. 2

    2020

  10. [18]

    Deepfakes dataset by google, jigsaw

    google. Deepfakes dataset by google, jigsaw. On- line, 2019. Available:https : / / blog . research . google / 2019 / 09 / contributing - data - to - deepfakedetection.html. 2

    2019

  11. [19]

    Spatiotemporal incon- sistency learning for deepfake video detection

    Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal incon- sistency learning for deepfake video detection. InProceed- ings of the 29th ACM international conference on multime- dia, pages 3473–3481, 2021. 2, 3

    2021

  12. [20]

    Spatiotemporal incon- sistency learning for deepfake video detection

    Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal incon- sistency learning for deepfake video detection. InACM Mul- timedia, pages 3473–3481, 2021. 2, 3, 4, 6, 7, 8

    2021

  13. [21]

    Hierarchical contrastive inconsistency learn- ing for deepfake video detection

    Zhihao Gu, Taiping Yao, Yang Chen, Shouhong Ding, and Lizhuang Ma. Hierarchical contrastive inconsistency learn- ing for deepfake video detection. InECCV, pages 596–613,

  14. [22]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 2

    2020

  15. [23]

    Owl- 1: Omni world model for consistent long video generation

    Yuanhui Huang, Wenzhao Zheng, Yuan Gao, Xin Tao, Pengfei Wan, Di Zhang, Jie Zhou, and Jiwen Lu. Owl- 1: Omni world model for consistent long video generation. arXiv preprint arXiv:2412.09600, 2024. 2, 3, 4

    2024 arXiv

  16. [24]

    The kinetics hu- 9 man action video dataset.arXiv preprint arXiv:1705.06950,

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics hu- 9 man action video dataset.arXiv preprint arXiv:1705.06950,

    arXiv

  17. [25]

    Auto-encoding vari- ational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding vari- ational bayes, 2013. 2

    2013

  18. [26]

    Video-llava: Learning united visual repre- sentation by alignment before projection

    Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual repre- sentation by alignment before projection. InProceedings of the 2024 Conference on Empirical Methods in Natural Lan- guage Processing, pages 5971–5984, 2024. 4

    2024

  19. [27]

    Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision

    Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pag...

    2024

  20. [28]

    Decof: Generated video detection via frame con- sistency: The first benchmark dataset.arXiv preprint arXiv:2402.02085, 2024

    Long Ma, Jiajia Zhang, Hongping Deng, Ningyu Zhang, Qinglang Guo, Haiyang Yu, Yong Liao, and Pengyuan Zhou. Decof: Generated video detection via frame con- sistency: The first benchmark dataset.arXiv preprint arXiv:2402.02085, 2024. 2, 6

    2024 arXiv

  21. [29]

    Latte: Latent diffusion transformer for video generation.arXiv preprint arXiv:2401.03048, 2024

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zi- wei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation.arXiv preprint arXiv:2401.03048, 2024. 2, 3, 4, 5

    2024 arXiv

  22. [30]

    Impacts and risk of generative ai tech- nology on cyber defense.arXiv preprint arXiv:2306.13033,

    Subash Neupane, Ivan A Fernandez, Sudip Mittal, and Shahram Rahimi. Impacts and risk of generative ai tech- nology on cyber defense.arXiv preprint arXiv:2306.13033,

    arXiv

  23. [31]

    Genvidbench: A challenging benchmark for detecting ai-generated video.arXiv preprint arXiv:2501.11340, 2025

    Zhenliang Ni, Qiangyu Yan, Mouxiao Huang, Tianning Yuan, Yehui Tang, Hailin Hu, Xinghao Chen, and Yunhe Wang. Genvidbench: A challenging benchmark for detecting ai-generated video.arXiv preprint arXiv:2501.11340, 2025. 2, 3, 5, 6, 8

    2025

  24. [32]

    Thinking in frequency: Face forgery detection by min- ing frequency-aware clues

    Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by min- ing frequency-aware clues. InECCV, pages 86–103, 2020. 2, 4, 6, 7, 8

    2020

  25. [33]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2

    2022

  26. [34]

    Faceforen- sics++: Learning to detect manipulated facial images

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Chris- tian Riess, Justus Thies, and Matthias Nießner. Faceforen- sics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1–11, 2019. 2

    2019

  27. [35]

    Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection

    Chuangchuang Tan, Huan Liu, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. InCVPR, 2024. 2, 4, 6, 7, 8

    2024

  28. [36]

    VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. InAdvances in Neural Information Processing Systems, 2022. 6, 7, 8

    2022

  29. [37]

    3d reconstruction with spatial memory.arXiv preprint arXiv:2408.16061, 2024

    Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory.arXiv preprint arXiv:2408.16061, 2024. 4, 5

    2024 arXiv

  30. [38]

    Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571, 2023

    Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571, 2023. 2, 3, 4, 5, 7, 8

    2023 arXiv

  31. [39]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697– 20709, 2024. 2, 4, 5

    2024

  32. [40]

    Drivedreamer: Towards real-world- drive world models for autonomous driving

    Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jia- gang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world- drive world models for autonomous driving. InEuropean Conference on Computer Vision, pages 55–72. Springer,

  33. [41]

    Lavie: High-quality video gener- ation with cascaded latent diffusion models.arXiv preprint arXiv:2309.15103, 2023

    Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video gener- ation with cascaded latent diffusion models.arXiv preprint arXiv:2309.15103, 2023. 3, 4, 5, 7, 8

    2023 arXiv

  34. [42]

    The emergence of deepfake technology: A review.Technology innovation management review, 9(11),

    Mika Westerlund. The emergence of deepfake technology: A review.Technology innovation management review, 9(11),

  35. [43]

    Pandora: Towards general world model with natural language actions and video states.arXiv preprint arXiv:2406.09455, 2024

    Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states.arXiv preprint arXiv:2406.09455, 2024. 2, 3

    2024 arXiv

  36. [44]

    Tall: Thumbnail layout for deepfake video detection

    Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. InProceedings of the IEEE/CVF interna- tional conference on computer vision, pages 22658–22668,

  37. [45]

    A survey on deepfake video detection.Iet Biometrics, 10(6): 607–624, 2021

    Peipeng Yu, Zhihua Xia, Jianwei Fei, and Yujiang Lu. A survey on deepfake video detection.Iet Biometrics, 10(6): 607–624, 2021. 2

    2021

  38. [46]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023. 3

    2023

  39. [47]

    Occworld: Learning a 3d occupancy world model for autonomous driving

    Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. InEuro- pean conference on computer vision, pages 55–72. Springer,

  40. [48]

    Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404, 2024

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404, 2024. 2, 3, 4, 5 10

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.