Pith. sign in

REVIEW 4 major objections 5 minor 59 references

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

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read SphereVideo claims that a single learnable real-region cap on a hypersphere can anchor a continual AI-video detector, preserving old knowledge and catching unseen generators.

desk verdict Solid but overclaimed: the method and benchmark are useful and the main tables hold up, but the 'unseen' generalization claim collapses to near-chance on the high-quality set in Supplement J. read the letter →

arxiv 2608.01334 v1 pith:ULFXPXTY submitted 2026-08-02 cs.CV

classification cs.CV
keywords AI-generatedvideodetectioncontinuallearninghypersphericalboundaryrealprototypecatastrophicforgettingtemporalcoherenceshufflecontrastopen-worldgeneralization
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 tries to establish that continual AI-generated-video detection can be stabilized by replacing hyperplane decision boundaries with a single "real region" on a unit hypersphere. It claims that real videos from different sources share a compact feature cluster after normalization, and that a learnable prototype $p_r$ with a learnable cosine margin $m$ defines a spherical cap: clips with $s=\langle \hat{x}_c,p_r\rangle > m$ are real, everything outside is fake. Because the anchor models "real" rather than any specific generator, the boundary resists drift when new fake generators arrive, and unseen generators fall outside the cap. The paper additionally claims that frame-level smoothness and clip-level shuffle-contrast losses on real clips make the real region more compact and improve temporal artifact sensitivity. If these claims are right, a detector can keep improving session by session on new generators while retaining accuracy on old ones, and open-world generalization to unseen generators improves by 4.00 accuracy points over prior continual methods.

What carries the argument

Prototype-anchored hyperspherical boundary: the clip feature is normalized, $\hat{x}_c = x_c/\|x_c\|_2$; a learnable unit vector $p_r$ is the real prototype; cosine similarity $s=\langle \hat{x}_c,p_r\rangle$; a learnable margin $m\in[0,1)$; logits $(s-m,\ m-s)\cdot\alpha$. The cap $\{s > m\}$ is the real region. It is trained by a spherical centripetal loss, frame-level temporal smoothness, clip-level shuffle contrast, plus replay and distillation for continual learning. The boundary carries the argument because it is regulated by a quantity inferred from real data—the prototype—rather than by the fake data currently being learned.

What would settle it

Collect real videos from domains never used in the nine sessions (archival footage, screen recordings, dashcam, heavy CGI), compute each clip's cosine similarity to the learned real prototype, and plot the distribution together with unseen fake generators' similarities. If a large fraction of genuine real clips fall below the learned margin or a large fraction of unseen fake clips fall above it, the single-compact-cluster premise is falsified; the paper's support for that premise is a qualitative visualization.

Watch

Extended reading notes

Core claim

Central claim: "real" is a region, not a label set. After $\ell_2$-normalization, a learnable unit prototype $p_r$ and margin $m$ define a spherical cap; $s=\langle \hat{x}_c,p_r\rangle>m$ means real. The cap anchors to the stable real distribution, so the prototype barely moves over nine sessions and unseen generators fall outside it. Temporal smoothness and shuffle-contrast losses make real features more compact, lifting detection of segment-shuffled real clips from 10.5% to 99.7%. The paper reports gains of 3.08% on seen data and 4.00% on unseen fake generators over prior continual baselines.

Load-bearing premise

The load-bearing premise is that all real videos—regardless of source, content, resolution, or motion—map to a single compact cluster around one prototype in normalized feature space, so that one fixed cap can separate every real clip from every fake one.

Editorial extensions

If this is right

  • If a stable cap is all that is needed, adding a new generator session does not require retraining from scratch or storing many old fake examples: the replay buffer can stay small and the prototype still pins down the real region.
  • Unseen generators can be caught without seeing them: because they land outside the learned cap, the open-world protocol reports 91.15% average accuracy on eight commercial unseen models versus the best baseline's 84.80%.
  • Temporal coherence becomes learnable from real videos alone: the shuffle-contrast signal needs no fake labels, and the ablation shows segment-shuffled real clips are detected as non-real in 99.70% of cases.
  • The boundary survives cheap degradations: because magnitude is discarded, mAA stays above 93% under H.264 CRF up to 45 and degrades gracefully under blur and resizing.
  • Ordering of sessions matters little: across five different session orders the paper reports mAA between 94.33% and 94.68%, so the anchor rather than the sequence is doing the stabilizing work.

Reading between the lines

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

  • A single real cap is also an open-set anomaly detector: a testable extension is to feed the model authentic but out-of-domain video (archival film, screen recordings, game captures) and see whether the real region is wide enough; if those fall outside, the compactness claim is domain-limited, not universal.
  • The shuffle-contrast loss is label-free and could be run on unlabeled real video before any session begins; this would probably initialize the prototype and margin better, and the paper does not report such a pretraining experiment.
  • The paper measures prototype stability only in the fixed feature space of its video encoder; whether the learned margin transfers to a new or fine-tuned encoder is untested, and re-calibration may be required.
  • The clearest next stress test is within-session paradigm changes: the reported gains concentrate after the session where the generative family changes, so the method may be most valuable exactly when generator diversity is high, and a benchmark with more paradigm switches would show how much of the 3.08% seen-data gain is attributable to the cap versus temporal losses.
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

4 major / 5 minor

Summary. The paper proposes SphereVideo, a continual learning framework for AI-generated video (AIGV) detection. It normalizes clip features to a unit hypersphere, learns a real prototype p_r and a similarity margin m (Eqs. 4-7), and adds frame-level temporal smoothness and clip-level shuffle contrastive losses (Eqs. 8-9) to model temporal coherence. The authors construct a nine-session continual-learning benchmark spanning UNet-based, DiT-based, and autoregressive generators, and compare against replay and non-replay baselines under two protocols. The headline claims are a 3.08% improvement on seen data and a 4.00% improvement on unseen AI-generated data, with the latter measured on the GenBuster-200K open-world benchmark.

Significance. If the empirical results are robust, the prototype-anchored hyperspherical boundary is a conceptually clean mechanism for stabilizing continual AIGV detection, and the proposed benchmark is a useful resource covering multiple generative paradigms. The method is simple, the loss design is well motivated, and the boundary-stability analysis in Fig. 5 is informative. However, the claims currently outrun the evidence: the unseen-data generalization result is benchmark-specific, the central compactness premise is only qualitatively supported, and the experimental section provides no error bars or released code/data. These issues must be addressed before the paper can be accepted.

major comments (4)
  1. [Section 4.3 / Supplement J] The abstract's '4.00% on unseen AI-generated data' is measured only on GenBuster-200K (Protocol 2). Supplement J reports a direct zero-shot evaluation on a curated high-quality set of 400 videos from SeedDance 2.0 and HappyHorse: SphereVideo achieves 54.75% accuracy, barely above chance, while the strongest baselines are at 40.75% and 33.00%. This is a load-bearing limitation. The abstract and conclusion should qualify the 'unseen AI-generated data' claim as benchmark-specific, or the method needs to be evaluated on high-quality AIGVs before claiming open-world generalization.
  2. [Section 3.4 / Eq. (7) / Supplement I] The entire method assumes that real video features are compact around a single prototype. Supplement I shows one qualitative UMAP from an external SUR-LID model, but provides no quantitative compactness measure. Please report the cosine-similarity distribution of real clips from each of the four real sources relative to the learned prototype p_r, compare it with the distribution for fake clips, and report what fraction of real features fall inside the learned hyperspherical cap. Without such evidence, the single-prototype 'real region' may misclassify diverse real content.
  3. [Tables 1-4 / Section 4.1] All experimental comparisons are single-run, with no standard deviations or confidence intervals. The benchmark is author-constructed and not released, and only a fixed random seed is mentioned. Since some gains are small (e.g., mAA 94.68 vs. 94.21 in Table 4), noise could change the conclusions. Please run at least three seeds and report means/std, and release the code, data splits, and benchmark construction details for reproducibility.
  4. [Section 3.5 / Table 4] The largest gains from the temporal losses are on perturbations generated by exactly the same operations used in the training losses: Acc_shuf and Acc_swap are measured on segment-shuffled and frame-swapped real videos, respectively. This is partially circular. The Protocol 1 mAA gain is only 0.47 points (Table 4), and Supplement J shows that TCLS improves high-quality zero-shot accuracy by 7.25 points but still leaves the model near chance. The claim that temporal coherence learning transfers to natural temporal artifacts of photorealistic unseen videos is therefore not supported by the current evidence.
minor comments (5)
  1. [Eq. (15)] All loss components are assigned equal weights. No sensitivity analysis is provided for alpha, tau, L, or replay-buffer size. A short ablation or a discussion of the chosen weights would help.
  2. [Table 4] The ordering of rows is a little confusing: 'SphereVideo (ours)' appears before 'w/ FTCN' and 'w/ AltFreezing'. Consider grouping baseline rows and then showing the final method.
  3. [Supplement B, Table 5] The formatting of the rows for Zeroscope/VideoCrafter1 and InternVid is garbled, making it hard to read the dataset statistics.
  4. [Fig. 5b] The UMAP plot has no axis values or legend for the ring-like structure, so it is difficult to assess how compact the real region actually is. Adding quantitative compactness statistics would be more informative.
  5. [Section 4.1] The phrase 'Zeroscope from DVF' is ambiguous because DVF refers to the benchmark paper [28]; please clarify that Zeroscope videos are generated using the DVF pipeline.

Circularity Check

1 steps flagged · score 3.0 of 10

Partial circularity in the temporal ablation (Tab. 4 Acc_shuf re-uses the training shuffle); the headline benchmark gains are empirical and independent, though Supp. J narrows the unseen-data claim.

  1. fitted input called prediction [Sec. 3.5, Eq. (9); Sec. 4.4, Table 4 and surrounding text]
    "we randomly partition each real clip into L consecutive temporal segments and then shuffle these segments to generate negative examples ... The contrastive loss L_shuffle is formulated as ...; "we generate purely temporal artifacts by reordering real data under two challenging configurations: 1) segment shuffling (consistent with our method) ... We define Acc_shuf ... to measure the accuracy of classifying such data as non-real.""

    Eq. (9) trains the model to push features of segment-shuffled real clips away from the original clip (equivalently, to treat shuffled clips as non-real). Table 4's Acc_shuf then evaluates the model on exactly the same segment-shuffling perturbation used to construct the training negatives. The reported 99.70% therefore largely measures whether the model recognizes its own augmentation, not an independent sensitivity to temporal incoherence. The adjacent-frame-swapping Acc_swap is less forced because that perturbation differs from the training shuffle, though it still aligns with the frame-level smoothness objective. This partial circularity affects only the Tab. 4 ablation claim; the central Protocol 1/2 benchmark comparisons in Tabs. 1–2 are independent empirical evaluations.

full rationale

The core derivation is not circular. The hyperspherical boundary (Eqs. 4–6), centripetal loss (Eq. 7), temporal coherence losses (Eqs. 8–9), and total loss (Eq. 15) define a concrete learning objective; classification performance is then measured against external baselines on the authors' benchmark (Tab. 1) and on GenBuster-200K (Tab. 2), with no fitted parameter being reported as a prediction. Fig. 1a's compactness observation comes from an external SUR-LID model, not from the proposed method. No load-bearing uniqueness theorem or self-citation chain is invoked. The one genuine circular step is Tab. 4's Acc_shuf: because the test perturbation is the same segment-shuffling operation used to generate negatives for Lshuffle, the 99.70% figure is partly the model recognizing its own training augmentation. The review rule also requires flagging the limitation in Supplement J: on the curated 400-video high-quality set (SeedDance 2.0 / HappyHorse), the full model scores only 54.75% zero-shot accuracy, while strong baselines are at 33.00–47.50%; this undercuts the abstract's unqualified '4.00% on unseen AI-generated data' claim, but it is a benchmark-specificity weakness rather than circularity. Overall, a moderate score of 3 reflects one partial self-referential ablation metric while the central empirical claims remain independent.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central method introduces two learned geometric quantities, the real prototype and the similarity margin, that define the decision boundary; neither is derived from first principles. The domain assumptions about real-data compactness and temporal smoothness are load-bearing and only qualitatively evidenced. No new physical entities are postulated.

free parameters (3)
  • Learnable real prototype p_r = unit vector in R^768 learned end-to-end
    Center of the real region; defines what "real" means and is optimized on the training and replay streams, so the boundary is data-fit rather than derived from first principles.
  • Learnable similarity margin m = initial approximately 0.497, final approximately 0.492
    Threshold radius of the real cap; trained with the model rather than set by theory, making the reported boundary stability an empirical property of a fitted scalar.
  • Training hyperparameters alpha, tau, L, replay configuration = alpha=20, tau=20, L=2, 64 replay samples per class, 4 replayed per class per batch, 5 epochs, lr 1e-5
    Chosen by hand; standard values rather than constants fitted to the reported metric, but they influence every comparison and ablation.
assumptions (5)
  • domain assumption Real video features, across YouTube-8M, InternVid, OpenVid, and HD-VG, form a single compact cluster around one prototype in normalized VideoMAE feature space, and every AI generator's features lie outside the induced cap.
    Sec. 1, Fig. 1a and Supplement I. This is the motivation for a one-prototype boundary; only qualitative UMAP support is provided.
  • domain assumption Only feature direction, not magnitude, is a reliable real/fake cue.
    Sec. 1, citing SphereFace [23] for face recognition and transferring the claim to AIGV without a dedicated video experiment.
  • domain assumption Adjacent real frames have smooth feature directions, and segment-shuffled real clips are non-real-like.
    Sec. 3.5, Eqs. (8)-(9). Real videos with cuts, fast motion, or camera changes can violate smoothness.
  • domain assumption Frozen-teacher distillation plus class-balanced replay preserves previous-session knowledge in VideoMAE.
    Sec. 3.6, Eqs. (12)-(14). Standard continual-learning assumption; the paper does not ablate the replay or distillation design independently of the overall results.
  • domain assumption VideoMAE pretrained on Kinetics-400 provides spatiotemporal features that reflect intrinsic real video structure.
    Sec. 3.3. The entire prototype and temporal geometry rests on this backbone's feature space.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SphereVideo: Prototype-anchored Hyperspherical Boundary for Continual AI-generated Video Detection." pith.science (2026). https://pith.science/paper/ULFXPXTY

@misc{pith2026260801334,
  author       = {Pith},
  title        = {Pith review of: SphereVideo: Prototype-anchored Hyperspherical Boundary for Continual AI-generated Video Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ULFXPXTY}},
  note         = {Machine review of arXiv:2608.01334}
}
read the original abstract

AI-generated video (AIGV) detection aims to distinguish real videos from AI-generated ones. In practice, detectors trained on existing data often fail to generalize to newly emerging generative models, making this task challenging. Therefore, continual learning (CL) is essential for improving the adaptability. However, CL frameworks for this task remain underexplored. To this end, we propose SphereVideo, a novel CL framework for AIGV detection built on two key observations. First, real videos exhibit a compact feature distribution. Based on this, we encourage real video features to cluster around a real prototype on a hypersphere while repelling AI-generated samples, thereby establishing a decision boundary. This prototype serves as a stable anchor for CL, regulating boundary evolution and mitigating catastrophic forgetting. Second, existing methods tend to rely solely on spatial artifacts as shortcuts. To enhance temporal modeling, we introduce a strategy that models the temporal dynamics of real data at both frame and clip levels. By strengthening real data modeling, this strategy further facilitates learning a real prototype and forming a stable decision boundary. Moreover, we construct a comprehensive and challenging benchmark. Extensive experiments demonstrate that SphereVideo achieves an improved plasticity-stability trade-off, outperforming prior methods by 3.08% on seen data and 4.00% on unseen AI-generated data.

Figures

Figures reproduced from arXiv: 2608.01334 by the authors.

Figure 2
Figure 2. Illustration of our second observation and corre [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of our SphereVideo framework. (a) Input samples are processed through a temporal encoder to extract frame￾level features, which are aggregated into clip-level feature to derive logits via hyperspherical head, followed by softmax-based prediction. Replay sample is used to preserve prior knowledge in new session. (b) The hyperspherical head derives logits via a Prototype-anchored Hyperspherical Boundary and e… view at source ↗
Figure 5
Figure 5. In-depth visualization. (a) Global and local similarities measure the cosine similarity of the current prototype p𝑟 against the initial (Session 0) and previous sessions, respectively. The simi￾larity margin 𝑚 is dynamically updated across sessions. (b) UMAP visualization of the feature distribution at the final session. of real data while maintaining a well-controlled margin, thereby establishing a robust decision … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Robustness evaluation of different methods. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Detailed performance of each method under protocol 1. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Visualization of boundary stability for our method [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 9
Figure 9. Figure 9: Grad-CAM visualization results of the proposed SphereVideo framework. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 31 canonical work pages

  1. [1]

    Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. 2016. Youtube-8m: A large-scale video classification benchmark.arXiv preprint arXiv:1609.08675 (2016)

  2. [2]

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

  3. [3]

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Leo Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. 2024. Video generation models as world simulators.OpenAI Blog1, 8 (2024), 1

  4. [4]

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

  5. [5]

    Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. 2023. Videocrafter1: Open diffusion models for high-quality video generation.arXiv preprint arXiv:2310.19512(2023)

  6. [6]

    Weiliang Chen, Wenzhao Zheng, Yu Zheng, Lei Chen, Jie Zhou, Jiwen Lu, and Yueqi Duan. 2025. Genworld: Towards detecting ai-generated real-world simula- tion videos.arXiv preprint arXiv:2506.10975(2025)

  7. [7]

    Jikang Cheng, Zhiyuan Yan, Ying Zhang, Li Hao, Jiaxin Ai, Qin Zou, Chen Li, and Zhongyuan Wang. 2025. Stacking brick by brick: Aligned feature isolation for incremental face forgery detection. InProceedings of the Computer Vision and Pattern Recognition Conference. 13927–13936

  8. [8]

    Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. 2024. Exploiting style latent flows for generalizing deepfake video detec- tion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1133–1143

Show all 59 references
  1. [9]

    Andrea Ciamarra, Roberto Caldelli, and Alberto Del Bimbo. 2024. Temporal surface frame anomalies for deepfake video detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 3837–3844

  2. [10]

    Riccardo Corvi, Davide Cozzolino, Ekta Prashnani, Shalini De Mello, Koki Nagano, and Luisa Verdoliva. [n. d.]. Seeing What Matters: Generalizable AI-generated Video Detection with Forensic-Oriented Augmentation. InThe Thirty-ninth An- nual Conference on Neural Information Proc...

  3. [11]

    Davide Cozzolino, Andreas Rössler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. 2021. Id-reveal: Identity-aware deepfake video detection. InProceedings of the IEEE/CVF international conference on computer vision. 15108–15117

  4. [12]

    Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. 2021. A continual learning survey: Defying forgetting in classification tasks.IEEE transactions on pattern analysis and machine intelligence44, 7 (2021...

  5. [13]

    Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Ting Zhang, Weiming Zhang, Neng- hai Yu, Dong Chen, Fang Wen, and Baining Guo. 2022. Protecting celebrities from deepfake with identity consistency transformer. InProceedings of the IEEE/CVF conference on computer vision and pattern rec...

  6. [14]

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. 2024. Ltx-video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103(2024)

  7. [15]

    Peisong He, Leyao Zhu, Jiaxing Li, Shiqi Wang, and Haoliang Li. 2024. Exposing ai-generated videos: A benchmark dataset and a local-and-global temporal defect based detection method.arXiv preprint arXiv:2405.04133(2024)

  8. [16]

    Yongkang Hu, Yu Cheng, Yushuo Zhang, Yuan Xie, and Zhaoxia Yin. 2025. SAIDO: Generalizable Detection of AI-Generated Images via Scene-Aware and Importance-Guided Dynamic Optimization in Continual Learning.arXiv preprint arXiv:2512.00539(2025)

  9. [17]

    Christian Internò, Robert Geirhos, Markus Olhofer, Sunny Liu, Barbara Hammer, and David Klindt. [n. d.]. AI-Generated Video Detection via Perceptual Straight- ening. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

  10. [18]

    Christian Internò, Robert Geirhos, Markus Olhofer, Sunny Liu, Barbara Hammer, and David Klindt. 2025. Ai-generated video detection via perceptual straightening. arXiv preprint arXiv:2507.00583(2025)

  11. [19]

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. 2024. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603 (2024)

  12. [20]

    Jiaming Li, Hongtao Xie, Jiahong Li, Zhongyuan Wang, and Yongdong Zhang

  13. [21]

    Zhengcen Li, Chenyang Jiang, Hang Zhao, Shiyang Zhou, Yunyang Mo, Feng Gao, Fan Yang, Qiben Shan, Shaocong Wu, and Jingyong Su. [n. d.]. Preserv- ing Forgery Artifacts: AI-Generated Video Detection at Native Scale. InThe Fourteenth International Conference on Learning Representations

  14. [22]

    Qingyuan Liu, Yun-Yun Tsai, Ruijian Zha, Victoria Li, Pengyuan Shi, Chengzhi Mao, and Junfeng Yang. 2025. Lavid: An agentic lvlm framework for diffusion- generated video detection.arXiv preprint arXiv:2502.14994(2025)

  15. [23]

    Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. 2017. Sphereface: Deep hypersphere embedding for face recognition. InProceedings of the IEEE conference on computer vision and pattern recognition. 212–220

  16. [24]

    Yue Ma, Kunyu Feng, Zhongyuan Hu, Xinyu Wang, Yucheng Wang, Mingzhe Zheng, Bingyuan Wang, Qinghe Wang, Xuanhua He, Hongfa Wang, et al. 2025. Controllable video generation: A survey.arXiv preprint arXiv:2507.16869(2025)

  17. [25]

    Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. 2024. Openvid-1m: A large-scale high-quality dataset for text-to-video generation.arXiv preprint arXiv:2407.02371(2024)

  18. [26]

    Zhenliang Ni, Qiangyu Yan, Mouxiao Huang, Tianning Yuan, Yehui Tang, Hailin Hu, Xinghao Chen, and Yunhe Wang. 2025. GenVidBench: A Challenging Bench- mark for Detecting AI-Generated Video.arXiv e-prints(2025), arXiv–2501

  19. [27]

    Kun Pan, Yifang Yin, Yao Wei, Feng Lin, Zhongjie Ba, Zhenguang Liu, Zhibo Wang, Lorenzo Cavallaro, and Kui Ren. 2023. Dfil: Deepfake incremental learning by exploiting domain-invariant forgery clues. InProceedings of the 31st ACM International Conference on Multimedia. 8035–8046

  20. [28]

    Xiufeng Song, Xiao Guo, Jiache Zhang, Qirui Li, Lei Bai, Xiaoming Liu, Guangtao Zhai, and Xiaohong Liu. 2024. On learning multi-modal forgery representation for diffusion generated video detection.Advances in Neural Information Processing Systems37 (2024), 122054–122077

  21. [29]

    Ke Sun, Shen Chen, Taiping Yao, Xiaoshuai Sun, Shouhong Ding, and Rongrong Ji. 2025. Continual face forgery detection via historical distribution preserving. International Journal of Computer Vision133, 3 (2025), 1067–1084

  22. [30]

    Shuai Tang, Peisong He, Haoliang Li, Wei Wang, Xinghao Jiang, and Yao Zhao

  23. [31]

    Hansi Teng, Hongyu Jia, Lei Sun, Lingzhi Li, Maolin Li, Mingqiu Tang, Shuai Han, Tianning Zhang, WQ Zhang, Weifeng Luo, et al. 2025. Magi-1: Autoregressive video generation at scale.arXiv preprint arXiv:2505.13211(2025)

  24. [32]

    Jiahe Tian, Cai Yu, Xi Wang, Peng Chen, Zihao Xiao, Jizhong Han, and Yesheng Chai. 2024. Dynamic mixed-prototype model for incremental deepfake detection. InProceedings of the 32nd ACM International Conference on Multimedia. 8129– 8138

  25. [33]

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. 2022. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems35 (2022), 10078–10093

  26. [34]

    Danial Samadi Vahdati, Tai D Nguyen, Aref Azizpour, and Matthew C Stamm

  27. [35]

    Hanyi Wang, Jun Lan, Yaoyu Kang, Huijia Zhu, Weiqiang Wang, Zhuosheng Zhang, and Shilin Wang. 2026. Generalizable and Adaptive Continual Learning Framework for AI-generated Image Detection.IEEE Transactions on Multimedia (2026)

  28. [36]

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. 2024. A comprehensive survey of continual learning: Theory, method and application.IEEE transactions on pattern analysis and machine intelligence46, 8 (2024), 5362–5383

  29. [37]

    Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. 2023. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation.(2023).arXiv preprint arXiv:2305.10874(2023)

  30. [38]

    Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. 2023. Internvid: A large-scale video-text dataset for multimodal understanding and generation.arXiv preprint arXiv:2307.06942(2023)

  31. [39]

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li

  32. [40]

    Haiquan Wen, Yiwei He, Zhenglin Huang, Tianxiao Li, Zihan Yu, Xingru Huang, Lu Qi, Baoyuan Wu, Xiangtai Li, and Guangliang Cheng. 2025. Busterx: Mllm- powered ai-generated video forgery detection and explanation.arXiv preprint arXiv:2505.12620(2025)

  33. [41]

    Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. 2024. Easyanimate: A high-performance long video genera- tion method based on transformer architecture.arXiv preprint arXiv:2405.18991 (2024)

  34. [42]

    Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He

  35. [43]

    Zhiyuan Yan, Yong Zhang, Xinhang Yuan, Siwei Lyu, and Baoyuan Wu. 2023. Deepfakebench: A comprehensive benchmark of deepfake detection.arXiv preprint arXiv:2307.01426(2023)

  36. [44]

    Zhiyuan Yan, Yandan Zhao, Shen Chen, Mingyi Guo, Xinghe Fu, Taiping Yao, Shouhong Ding, Yunsheng Wu, and Li Yuan. 2025. Generalizing deepfake video detection with plug-and-play: Video-level blending and spatiotemporal adapter tuning. InProceedings of the Computer Vision and Pa...

  37. [45]

    Chin-Chia Yang, Yung-Yu Chuang, Hwann-Tzong Chen, and Tyng-Luh Liu. [n. d.]. HSIC Bottleneck for Cross-Generator and Domain-Incremental Synthetic Image Detection. InThe Fourteenth International Conference on Learning Representations

  38. [46]

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al . 2024. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072(2024)

  39. [47]

    Tall: Thumbnail layout for deepfake video detection. InProceedings of the MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Fei Li, Yue Yu, Yuran Wang, Xinghan Li, Jingjing Chen, and Yu-Gang Jiang IEEE/CVF international conference on computer vision. 22658–22668

  40. [48]

    Daichi Zhang, Zihao Xiao, Shikun Li, Fanzhao Lin, Jianmin Li, and Shiming Ge. 2024. Learning natural consistency representation for face forgery video detection. InEuropean conference on computer vision. Springer, 407–424

  41. [49]

    Shuhai Zhang, ZiHao Lian, Jiahao Yang, Daiyuan Li, Guoxuan Pang, Feng Liu, Bo Han, Shutao Li, and Mingkui Tan. [n. d.]. Physics-Driven Spatiotemporal Model- ing for AI-Generated Video Detection. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

  42. [50]

    Tianshuo Zhang, Li Gao, Siran Peng, Xiangyu Zhu, and Zhen Lei. 2025. DevFD: Developmental Face Forgery Detection by Learning Shared and Orthogonal LoRA Subspaces. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

  43. [51]

    Xueyi Zhang, Peiyin Zhu, Chengwei Zhang, Zhiyuan Yan, Jikang Cheng, Min- grui Lao, Siqi Cai, and Yanming Guo. 2025. Generalization-Preserved Learning: Closing the Backdoor to Catastrophic Forgetting in Continual Deepfake Detec- tion. InProceedings of the IEEE/CVF International...

  44. [52]

    Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. 2025. From slow bidirectional to fast autoregressive video diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22963–22974

  45. [53]

    Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. 2021. Ex- ploring temporal coherence for more general video face forgery detection. In Proceedings of the IEEE/CVF international conference on computer vision. 15044– 15054

  46. [54]

    Qty. ” denotes the number of videos, and “Dur

    Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, Lijun Zhang, and De-Chuan Zhan. 2025. Dual consolidation for pre-trained model-based domain-incremental learning. In Proceedings of the Computer Vision and Pattern Recognition Conference. 20547– 20557. SphereVideo: Prototype-anchored Hypers...

  47. [57]

    Chende Zheng, Ruiqi Suo, Chenhao Lin, Zhengyu Zhao, Le Yang, Shuai Liu, Minghui Yang, Cong Wang, and Chao Shen. 2025. D3: Training-free ai-generated video detection using second-order features. InProceedings of the IEEE/CVF International Conference on Computer Vision. 12852–12862

  48. [2021]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Frequency-aware discriminative feature learning supervised by single- center loss for face forgery detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6458–6467

  49. [2023]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Altfreezing for more general video face forgery detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4129–4138

  50. [2024]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Beyond deepfake images: Detecting ai-generated videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 4397–4408

  51. [2025]

    Towards extensible detection of AI-generated images via content-agnostic adapter-based category-aware incremental learning.IEEE Transactions on Infor- mation Forensics and Security(2025)

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.