Pith. sign in

REVIEW 4 major objections 5 minor 62 references

HumanSAM: Classifying Human-centric Forgery Videos in Human Spatial, Appearance, and Motion Anomaly

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

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

Pith's one-line read This paper argues that human-centric video forgeries decompose into three named anomaly classes and that a dual-branch model can classify them, along with real videos, in one four-way task.

desk verdict New task and benchmark, but the multi-class evaluation is circular: labels come from the same Depth Pro/CLIP/RAFT pipeline the model consumes, so the headline numbers only measure self-agreement. read the letter →

arxiv 2507.19924 v2 pith:WRLFQQQ7 submitted 2025-07-26 cs.CV

Chang Liu , Yunfan Ye , Fan Zhang , Qingyang Zhou , Yuchuan Luo , Zhiping Cai This is my paper
classification cs.CV
keywords human-centricforgeryvideodetectionfine-grainedclassificationspatialanomalyappearancemotionHFVdatasetdual-branchfusionrank-basedconfidenceenhancement
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

This paper argues that the artifacts left by today's human-centric video generators are not one homogeneous "fake" signal but fall into three recognizable classes: spatial anomalies (geometric impossibilities), appearance anomalies (identity and object inconsistencies across frames), and motion anomalies (biomechanically unnatural movement). To make that claim usable, it introduces HFV, the first public benchmark built for this four-way classification (fake-spatial, fake-appearance, fake-motion, real), and HumanSAM, a model that fuses a video-understanding backbone with a frozen monocular depth branch and a rank-based loss weighting. The authors report that HumanSAM outperforms prior methods on both binary real-vs-fake and fine-grained multi-class classification. The point of the paper is that fine-grained forgery labeling is possible and improves both reliability and interpretability over binary detection.

What carries the argument

Human Forgery Representation (HFR), a dual-branch feature: a trainable video-understanding branch (InternVideo2) produces attention-pooled and average-pooled spatiotemporal tokens, while a frozen spatial-depth branch (Depth Pro) encodes the first and middle frames, and a learnable scalar $\alpha$ weights the fusion before linear projection to four classes. The rank-based confidence enhancement uses each video's normalized anomaly rank $\hat{r}_i = r_i/n$ from the labeling pipeline to weight the loss as $\log(e+\hat{r}_i)\,L_i$, up-weighting the most anomalous samples. The HFV labels themselves are generated by ranking three scores — depth-frame warping error (Depth Pro plus RAFT), sliding-window CLIP/DINOv2 similarity, and optical-flow warping error — and taking the strongest rank as the class.

What would settle it

Take a random sample of HFV videos, have several human annotators independently assign the dominant anomaly (spatial, appearance, motion), and measure agreement between those annotations and the automatic Depth-Pro/CLIP-DINOv2/RAFT label; if agreement is close to chance, the benchmark's reported accuracies would not reflect human-recognizable forgery types.

Watch

Extended reading notes

Core claim

The central claim is that human-centric forgeries can be decomposed into spatial, appearance, and motion anomaly, and that this decomposition is learnable from a representation that combines spatiotemporal semantic features with spatial-depth features. On the HFV benchmark, which labels videos by the strongest of three automatic anomaly scores, HumanSAM reaches 69.8% average multi-class accuracy and 87.9% AUC, and 97.8% accuracy and 99.9% AUC on the binary real-vs-fake task, beating state-of-the-art baselines in both settings. The authors also show that each design choice — the attention/average pooling fusion in the video branch, the depth branch, and the rank-based confidence loss — contributes a stepwise gain in the ablations.

Load-bearing premise

The load-bearing premise is that the automatic anomaly scores used to produce HFV's ground-truth labels match what a human would call the dominant flaw; if that fails, the reported accuracy measures agreement with an algorithmic oracle, not perceptual validity.

Editorial extensions

If this is right

  • A single detector can name the dominant flaw type of a generated human video, not just label it fake, which gives operators an interpretable reason for a flag.
  • Training on the three-way anomaly task transfers back to binary detection: on HFV it beats prior detectors in the real-vs-fake setting.
  • The benchmark exposes per-source difficulty; results drop sharply on Kling, suggesting that newer or higher-quality generators leave subtler anomalies that current detectors underuse.
  • Because labels come from scores that can be computed on any new generator, the same pipeline can be re-run to keep the benchmark current.

Reading between the lines

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

  • The paper leaves open whether the three anomaly classes are perceptually distinct to human viewers; a human-annotation study of the HFV labels would test that directly.
  • If the three-way decomposition is right, the same scoring pipeline could be used not only for detection but as a diagnostic for video-generation models, telling a developer which of the three failure modes to fix.
  • The method's accuracy is likely tied to the overlap between the frozen depth branch and the label pipeline's depth score; replacing Depth Pro in either place could change the measured performance.
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 HumanSAM, a framework for fine-grained classification of human-centric forgery videos into four classes: spatial, appearance, and motion anomaly, and real videos. The authors introduce the HFV dataset, where each video is automatically labeled by ranking three anomaly scores derived from pre-trained models: Depth Pro for spatial, CLIP/DINOv2 for appearance, and RAFT-based warping error for motion. The model itself uses a frozen Depth Pro branch for spatial features, a video understanding branch (InternVideo2), and a rank-based confidence loss that consumes the same anomaly ranks used to generate the labels. Experiments report state-of-the-art multi-class accuracy (69.8%) and AUC (87.9%) on HFV, along with ablations and robustness analyses.

Significance. If the benchmark and evaluations were valid, this would be a valuable step toward interpretable, fine-grained forgery detection in human-centric videos. The task formulation is sensible and the method design (dual-branch fusion with depth features, rank-based loss) is plausible. The paper also includes useful ablations and per-generator generalization experiments. However, the central contribution—the HFV benchmark—is currently undermined by a circular evaluation: the ground-truth labels are produced by the same scoring pipeline whose feature extractors (Depth Pro, CLIP/DINOv2, RAFT) are directly consumed by the proposed model, and the test-set labels come from the same pipeline without independent human verification. As a result, the reported performance does not establish that the model recognizes human-perceived anomaly types; it may simply reflect agreement with an algorithmic oracle. This issue is load-bearing because the paper's main claim is the SOTA performance on HFV and the utility of the benchmark.

major comments (4)
  1. [§3.2.2, Eq. (2)] The ground-truth labels in HFV are assigned by taking the highest rank among three anomaly scores (Eq. 2), where these scores are computed from Depth Pro for spatial, CLIP/DINOv2 for appearance, and RAFT-based warping error for motion. The proposed HumanSAM then uses a frozen Depth Pro branch (§3.3) and a rank-based confidence loss (§3.4) that directly consumes the same ranks used to define the labels. This creates a circular evaluation: the model is trained and tested against labels generated by the very same feature extractors and scores that the model uses. Consequently, the reported multi-class ACC/AUC (e.g., 69.8%/87.9% in Table 2) may largely measure the model's ability to reproduce the algorithmic scoring pipeline, rather than its ability to recognize perceptually meaningful spatial, appearance, or motion anomalies. The paper does not provide a human-annotated test set, inter-annotator agreement metrics, or any evidence that the algorithmic labels align with human judgments. This is the central weakness and should be addressed before the benchmark or the reported results can be considered valid.
  2. [Table 2 and §4.2] The comparison with existing detectors is biased by the circular label-generation process. Baseline methods (e.g., TimeSformer, MM-Det, Uni-FD) do not use Depth Pro features, yet the spatial-anomaly label is derived precisely from Depth Pro-based scores. As a result, the proposed method has an inherent advantage in predicting the spatial anomaly class, since it shares the exact feature extractor used to create the label. This makes the claimed 'SOTA' improvement in §4.2 uninterpretable as a measure of general detection ability. A fair comparison would require either a human-annotated test set with labels independent of any particular feature extractor, or at least a controlled setting where baselines are also given access to the same auxiliary features.
  3. [§3.2.3 and §4.1] The paper states that the top 20% of videos per anomaly class were 'supplemented by human review' to form the validation set, but it is not described what this review involved, how many annotators participated, or whether any quantitative agreement (e.g., Cohen's kappa) was measured. More importantly, the training set (the remaining 80%) and the test sets from the eight other generation models are not mentioned as being human-verified at all. The validation set is used for model selection, but the final reported results are on test sets whose labels are purely algorithmic. This is insufficient to break the circularity identified above. The authors should provide a human-annotated test subset with agreement statistics and show that model predictions align with human judgments, not just with the scoring pipeline.
  4. [§3.4, Eqs. (6)–(9)] The rank-based confidence enhancement loss uses the same per-class rankings that define the ground-truth labels. While the motivation (to emphasize high-confidence samples) is reasonable, the practical effect is to further entangle the training objective with the label-generation pipeline. Since the ranks are derived from the same three scores that also determine the label, the loss effectively teaches the model to be more confident about samples that the scoring pipeline considers 'obvious' anomalies. This does not provide independent evidence of robustness; it only amplifies the model's sensitivity to the pipeline's own judgments. The contribution of this component would be more convincing if the confidence weights were derived from an independent source or if the labels were human-verified.
minor comments (5)
  1. [§3.2.2] The text contains an inconsistency: it first says 'higher rank indicates greater anomaly severity' and then says 'lower rankings suggest the corresponding anomaly is less pronounced,' and Eq. (2) uses the smallest rank as the label. Please clarify the ranking convention consistently.
  2. [§3.4] There is a typo in the explanation: 'samples with with higher rankings' should be 'samples with higher rankings.'
  3. [§3.3, Eq. (5)] The dynamic fusion weight α is described as a learnable parameter, but no specific initialization or constraints are mentioned. Please provide details on how α is initialized and whether it is bounded during training.
  4. [Table 5] The robustness table reports AUC only, not ACC, and 'JPEG Compression (C) with Q=90' is described in the text, but the table header says 'JPEG Q=50'. Please ensure the compression quality is stated consistently.
  5. [§D.1 in Supplementary] The F1-scores for appearance and spatial anomalies are low (0.55 and 0.49), and the authors note that most confusion occurs between motion and spatial classes. This further suggests that the label categories may not be well-separated perceptually, which reinforces the need for human validation.

Circularity Check

2 steps flagged · score 8.0 of 10

HFV labels are generated by the same Depth Pro/CLIP/DINOv2/RAFT scoring pipeline that HumanSAM consumes, so the reported multi-class accuracy largely measures agreement with the auto-labeling oracle rather than an independent ground truth.

  1. self definitional [Sec. 3.2.2, Eq. (2); Sec. 3.3 'Spatial Depth Branch']
    "L(vi) = arg min c∈{s,a,m} Rc(vi) (2) where c ∈ {s, a, m} corresponds to spatial, appearance, and motion anomaly, respectively."

    The benchmark label L(vi) is defined as the argmin of ranks of the three anomaly scores; the spatial score is computed with Depth Pro plus RAFT warping error, the appearance score with CLIP/DINOv2 similarity, and the motion score with RAFT distortion. HumanSAM's Spatial Depth Branch is built from the same frozen Depth Pro encoder used to produce the spatial score. Since the eight evaluation sets are labeled by the same automatic ranking pipeline and no human verification is described for them, the reported multi-class ACC measures how well the model reproduces the auto-scoring oracle's argmax, not an independently verified perceptual anomaly judgment.

  2. fitted input called prediction [Sec. 3.4, Eqs. (6)-(9)]
    "Lweighted i = αi · Li = log(e + r̂i) · Li (8) ... Ltotal = 1/N Σ log(e + r̂i) · Li (9)"

    The rank r_i used to weight each training loss is exactly the rank R(v_i) computed during dataset construction, the same rank that determines the ground-truth label L(v_i) via Eq. (2). Thus the training objective is reweighted by the very scores that define the benchmark labels. The model is therefore explicitly optimized toward the same algorithmic scoring pipeline that later generates the test labels, so the reported fine-grained accuracy is not independent of the training target; it largely reflects agreement with that pipeline.

full rationale

The paper's central claim is fine-grained state-of-the-art classification on the new HFV benchmark. That claim depends on the benchmark labels being valid ground truth for spatial, appearance, and motion anomalies. But the paper itself defines the labels as the argmin over ranks of anomaly scores computed by Depth Pro + RAFT, CLIP/DINOv2, and RAFT distortion. HumanSAM then uses a frozen Depth Pro branch as one of its two feature sources, and its rank-based confidence loss reweights training by exactly the ranks that created the labels. The only human review described is for the validation split; the eight cross-model evaluation sets inherit the automatic labels. Consequently, the reported 69.8% multi-class ACC primarily measures how well the model can rediscover the auto-scoring pipeline's ranking, not whether the categories correspond to human-perceived anomaly types. This is a partial but central circularity: the benchmark ground truth and the model are not independent. There is no self-citation chain or uniqueness-theorem issue; the circularity is in the dataset/evaluation construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the automatic scoring pipeline being a valid ground-truth generator, a taxonomy assumption, and several hand-chosen parameters in the label construction; the model itself adds one learned fusion weight. No new physical entities are invoked.

free parameters (5)
  • fusion weight alpha
    Eq. 5: a learnable scalar blending the VU branch and SD branch features, trained on the HFV training set; the paper reports no final value.
  • appearance score weights alpha, beta = 0.5, 0.5
    Eq. 13: hand-set weights for the two terms of the appearance consistency score, directly controlling the appearance labels in the benchmark.
  • sliding window size W = 5
    Eq. 14: hand-chosen window length for temporal consistency similarity in the appearance scoring pipeline.
  • validation top-20% threshold = 20%
    Sec. 3.2.3: videos with the most obvious anomaly scores are selected for the validation set; the cutoff is arbitrary.
  • rank-to-confidence mapping = log(e + r_hat)
    Eq. 7: the mapping from normalized rank to loss weight is a hand-picked function; while it has no learned parameters, it is an ad hoc choice shaping training.
assumptions (4)
  • domain assumption The three anomaly types (spatial, appearance, motion) are exhaustive and mutually exclusive for human-centric forgeries.
    Sec. 1 and Sec. 3.2.2 assert this taxonomy based on observations and prior literature; no proof is given that all generated human videos fall into exactly one of these categories.
  • domain assumption Depth Pro monocular depth distortion, CLIP/DINOv2 cosine similarity, and RAFT optical-flow warping error are valid proxies for spatial, appearance, and motion anomalies respectively.
    Sec. 3.2.2 and Appendix A define the scores without comparing them to human judgments; the benchmark labels inherit this assumption.
  • ad hoc to paper Labeling a video by the highest relative rank among the three anomaly scores (Eq. 2) selects the most salient anomaly.
    The arg-min rule over ranks is introduced for this dataset; it may mislabel videos with multiple severe anomalies (the paper says <2% have all three) and gives no uncertainty estimate.
  • domain assumption The set of real videos from Kinetics-400 is an uncontaminated 'real' class.
    Sec. 3.2.1 uses K400 as real human action videos; no checks for synthetic or manipulated content are described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HumanSAM: Classifying Human-centric Forgery Videos in Human Spatial, Appearance, and Motion Anomaly." pith.science (2026). https://pith.science/paper/WRLFQQQ7

@misc{pith2026250719924,
  author       = {Pith},
  title        = {Pith review of: HumanSAM: Classifying Human-centric Forgery Videos in Human Spatial, Appearance, and Motion Anomaly},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WRLFQQQ7}},
  note         = {Machine review of arXiv:2507.19924}
}
read the original abstract

Numerous synthesized videos from generative models, especially human-centric ones that simulate realistic human actions, pose significant threats to human information security and authenticity. While progress has been made in binary forgery video detection, the lack of fine-grained understanding of forgery types raises concerns regarding both reliability and interpretability, which are critical for real-world applications. To address this limitation, we propose HumanSAM, a new framework that builds upon the fundamental challenges of video generation models. Specifically, HumanSAM aims to classify human-centric forgeries into three distinct types of artifacts commonly observed in generated content: spatial, appearance, and motion anomaly. To better capture the features of geometry, semantics and spatiotemporal consistency, we propose to generate the human forgery representation by fusing two branches of video understanding and spatial depth. We also adopt a rank-based confidence enhancement strategy during the training process to learn more robust representation by introducing three prior scores. For training and evaluation, we construct the first public benchmark, the Human-centric Forgery Video (HFV) dataset, with all types of forgeries carefully annotated semi-automatically. In our experiments, HumanSAM yields promising results in comparison with state-of-the-art methods, both in binary and multi-class forgery classification.

Figures

Figures reproduced from arXiv: 2507.19924 by the authors.

Figure 1
Figure 1. Vanilla video classifier distinguishes only fake or real [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The evolution process of pseudo-label generation. Videos are processed through an anomaly scoring mechanism to produce three scores: spatial, appearance, and motion anomalies. Based on these scores, videos are ranked by anomaly type, with the highest-ranked anomaly determining the final label. During dataset construction, videos with the same anomaly label are re-ranked, and the top 20% are reviewed for strong anoma… view at source ↗
Figure 3
Figure 3. Visual illustration of spatial anomaly in a generated video sequence. Consecutive frames from a synthetic video (top) and their depth maps (bottom) reveal inconsistencies—e.g., unnat￾ural clipping between hand and drum—despite visual appearance remaining similar. is a typical example of a spatial anomaly. Appearance anomaly. This anomaly involves the failure to maintain a consistent appearance of characters or objec… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Dual-branch framework for human-centric video forgery detection. The model comprises a spatio-temporal video under￾standing (VU) branch and a frozen spatial-depth (SD) branch. The VU branch encodes video frames using a transformer-based encoder to extract attention- an…
Figure 5
Figure 5. Figure 5: Confusion matrices on CogVideoX-5B dataset: (a) Multi-class, (b) Binary. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Spatial and temporal activation visualization on a CogVideoX-5B video. From the last Transformer layer of the video branch: [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: It can be seen that in the first row, two metal knives blur and pass through each other; in the second row, a woman’s hand blurs [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Some examples of appearance anomalies.It can be seen that in the first row, the hand on the right suddenly changes from an [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Some examples of motion anomalies.It can be observed that in the first row, the woman’s body maintains a forward-leaning [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 41 canonical work pages

  1. [1]

    https://vchitect.intern-ai.org.cn, 2024

    Vchitect 2.0. https://vchitect.intern-ai.org.cn, 2024. 3, 8

    work page 2024

  2. [2]

    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 The 7th Chinese Conference on Pattern Recognition and Computer Vision (PRCV), 2024. 1, 2, 3

    work page 2024

  3. [3]

    Is space-time attention all you need for video understanding? InICML, page 4, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? InICML, page 4, 2021. 3, 6, 7, 8, 13

    work page 2021

  4. [4]

    Depth pro: Sharp monocular metric depth in less than a second.arXiv preprint arXiv:2410.02073, 2024

    Aleksei Bochkovskii, Ama ¨el Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second.arXiv preprint arXiv:2410.02073, 2024. 3, 4, 6, 8, 12, 13

    arXiv 2024

  5. [5]

    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

  6. [6]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. InPro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 12

    work page 2021

  7. [7]

    What matters in detecting ai-generated videos like sora? arXiv preprint arXiv:2406.19568, 2024

    Chirui Chang, Zhengzhe Liu, Xiaoyang Lyu, and Xiaojuan Qi. What matters in detecting ai-generated videos like sora? arXiv preprint arXiv:2406.19568, 2024. 1, 2, 4

    arXiv 2024

  8. [8]

    Multi-view clustering via deep concept factorization

    Shuai Chang, Jie Hu, Tianrui Li, Hao Wang, and Bo Peng. Multi-view clustering via deep concept factorization. Knowledge-Based Systems, 217:106807, 2021. 2

    work page 2021

Show all 62 references
  1. [9]

    Videojam: Joint appearance-motion representations for en- hanced motion generation in video models.arXiv preprint arXiv:2502.02492, 2025

    Hila Chefer, Uriel Singer, Amit Zohar, Yuval Kirstain, Adam Polyak, Yaniv Taigman, Lior Wolf, and Shelly Sheynin. Videojam: Joint appearance-motion representations for en- hanced motion generation in video models.arXiv preprint arXiv:2502.02492, 2025. 2

    2025 arXiv

  2. [10]

    https://research.runwayml.com/gen2, 2023

    Runway Gen-2. https://research.runwayml.com/gen2, 2023. 3, 8

    2023

  3. [11]

    https://runwayml.com/research/introducing- gen-3-alpha, 2024

    Runway Gen-3. https://runwayml.com/research/introducing- gen-3-alpha, 2024. 1, 2, 3, 7, 8

    2024

  4. [12]

    Hierarchical fine-grained im- age forgery detection and localization

    Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Ia- copo Masi, and Xiaoming Liu. Hierarchical fine-grained im- age forgery detection and localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3155–3165, 2023. 1, 8

    2023

  5. [13]

    Venhancer: Generative space-time enhancement for video generation.arXiv preprint arXiv:2407.07667, 2024

    Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation.arXiv preprint arXiv:2407.07667, 2024. 3, 8

    2024 arXiv

  6. [14]

    Video dif- fusion models.Advances in Neural Information Processing Systems, 35:8633–8646, 2022

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video dif- fusion models.Advances in Neural Information Processing Systems, 35:8633–8646, 2022. 2

    2022

  7. [15]

    Vbench: Comprehensive bench- mark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive bench- mark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pa...

    2024

  8. [16]

    Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

    2001 arXiv

  9. [17]

    Large-scale video classification with convolutional neural networks

    Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. InPro- ceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725–1732, 2014. 3

    2014

  10. [18]

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

    arXiv

  11. [19]

    https://klingai.kuaishou.com/, 2024

    klingai. https://klingai.kuaishou.com/, 2024. 1, 2, 3, 14

    2024

  12. [20]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024. 1, 2

    2024 arXiv

  13. [21]

    https://pika.art, 2024

    Pika Labs. https://pika.art, 2024. 3

    2024

  14. [22]

    Learning blind video temporal consistency

    Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. InProceedings of the European con- ference on computer vision (ECCV), pages 170–185, 2018. 4, 12

    2018

  15. [23]

    Blind video temporal consistency via deep video prior.Advances in Neu- ral Information Processing Systems, 33:1083–1093, 2020

    Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior.Advances in Neu- ral Information Processing Systems, 33:1083–1093, 2020. 4, 12

    2020

  16. [24]

    A comprehensive survey on human video gener- ation: Challenges, methods, and insights.arXiv preprint arXiv:2407.08428, 2024

    Wentao Lei, Jinting Wang, Fengji Ma, Guanjie Huang, and Li Liu. A comprehensive survey on human video gener- ation: Challenges, methods, and insights.arXiv preprint arXiv:2407.08428, 2024. 4

    2024 arXiv

  17. [25]

    Unmasked teacher: Towards training-efficient video foundation models

    Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 19948–19960, 2023. 3

    2023

  18. [26]

    Evalcrafter: Benchmarking and eval- uating large video generation models

    Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and eval- uating large video generation models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

    2024

  19. [27]

    Step-video-t2v technical re- port: The practice, challenges, and future of video founda- tion model.arXiv preprint arXiv:2502.10248, 2025

    Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xi- aoniu Song, Xing Chen, et al. Step-video-t2v technical re- port: The practice, challenges, and future of video founda- tion model.arXiv preprint arXiv:2502.10248, 2025. 1, 2

    2025 arXiv

  20. [28]

    https://hailuoai.com/video, 2024

    MiniMax. https://hailuoai.com/video, 2024. 1, 2, 3, 7, 14 9

    2024

  21. [29]

    Towards uni- versal fake image detectors that generalize across genera- tive models

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards uni- versal fake image detectors that generalize across genera- tive models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480– 24489, 2023. 2, 6, 7, 8, 13

    2023

  22. [30]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4, 12

    2023 arXiv

  23. [31]

    Fine-grained bipar- tite concept factorization for clustering

    Chong Peng, Pengfei Zhang, Yongyong Chen, Zhao Kang, Chenglizhao Chen, and Qiang Cheng. Fine-grained bipar- tite concept factorization for clustering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26264–26274, 2024. 2

    2024

  24. [32]

    Fatezero: Fus- ing attentions for zero-shot text-based video editing

    Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fus- ing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932–15942, 2023. 4, 12

    2023

  25. [33]

    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. InEuropean conference on com- puter vision, pages 86–103. Springer, 2020. 1, 2, 6, 7, 8, 13

    2020

  26. [34]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, p...

    2021

  27. [35]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024. 3

    2024 arXiv

  28. [36]

    De-fake: Detection and attribution of fake images generated by text- to-image generation models

    Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. De-fake: Detection and attribution of fake images generated by text- to-image generation models. InProceedings of the 2023 ACM SIGSAC Conference on Computer and Communica- tions Security, pages 3418–3432, 2023. 1, 2

    2023

  29. [37]

    Repvideo: Rethinking cross- layer representation for video generation.arXiv preprint arXiv:2501.08994, 2025

    Chenyang Si, Weichen Fan, Zhengyao Lv, Ziqi Huang, Yu Qiao, and Ziwei Liu. Repvideo: Rethinking cross- layer representation for video generation.arXiv preprint arXiv:2501.08994, 2025. 2

    2025 arXiv

  30. [38]

    Two-stream con- volutional networks for action recognition in videos.Ad- vances in neural information processing systems, 27, 2014

    Karen Simonyan and Andrew Zisserman. Two-stream con- volutional networks for action recognition in videos.Ad- vances in neural information processing systems, 27, 2014. 2, 3

    2014

  31. [39]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 7

    2010 arXiv

  32. [40]

    On learn- ing multi-modal forgery representation for diffusion gener- ated video detection

    Xiufeng Song, Xiao Guo, Xiaohong Liu, Jiache Zhang, Qirui Li, Lei Bai, Xiaoming Liu, and Guangtao Zhai. On learn- ing multi-modal forgery representation for diffusion gener- ated video detection. InProceeding of Thirty-eighth Confer- ence on Neural Information Processing Syste...

    2024

  33. [41]

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

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling op- erations in cnn-based generative network for generalizable deepfake detection. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, ...

    2024

  34. [42]

    Raft: Recurrent all-pairs field transforms for optical flow

    Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23– 28, 2020, Proceedings, Part II 16, pages 402–419. Springer,

    2020

  35. [43]

    Learning spatiotemporal features with 3d convolutional networks

    Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. InProceedings of the IEEE inter- national conference on computer vision, pages 4489–4497,

  36. [44]

    https://wanxai.com/, 2025

    Wan2.1. https://wanxai.com/, 2025. 1, 2

    2025

  37. [45]

    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

    2023 arXiv

  38. [46]

    Cnn-generated images are surprisingly easy to spot

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695–8704, 2020. 1, 2, 6, 7, 8, 13

    2020

  39. [47]

    Internvideo2: Scaling video foundation mod- els for multimodal video understanding.arXiv preprint arXiv:2403.15377, 2024

    Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation mod- els for multimodal video understanding.arXiv preprint arXiv:2403.15377, 2024. 3, 5, 8, 13

    2024 arXiv

  40. [48]

    Dire for diffusion-generated image detection

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445–22455, 2023. 1, 2, 6, 7, 8, 13

    2023

  41. [49]

    Finepose: Fine- grained prompt-driven 3d human pose estimation via diffu- sion models

    Jinglin Xu, Yijie Guo, and Yuxin Peng. Finepose: Fine- grained prompt-driven 3d human pose estimation via diffu- sion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 561–570, 2024. 3

    2024

  42. [50]

    Fineparser: A fine-grained spatio-temporal action parser for human-centric action quality assessment

    Jinglin Xu, Sibo Yin, Guohao Zhao, Zishuo Wang, and Yuxin Peng. Fineparser: A fine-grained spatio-temporal action parser for human-centric action quality assessment. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14628–14637...

    2024

  43. [51]

    Human motion video genera- tion: A survey.Authorea Preprints, 2024

    Haiwei Xue, Xiangyang Luo, Zhanghao Hu, Xin Zhang, Xunzhi Xiang, Yuqin Dai, Jianzhuang Liu, Zhensong Zhang, Minglei Li, Jian Yang, et al. Human motion video genera- tion: A survey.Authorea Preprints, 2024. 4

    2024

  44. [52]

    Thinking in space: How mul- timodal large language models see, remember, and recall spaces.arXiv preprint arXiv:2412.14171, 2024

    Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in space: How mul- timodal large language models see, remember, and recall spaces.arXiv preprint arXiv:2412.14171, 2024. 3

    2024 arXiv

  45. [53]

    Depth any- thing v2.arXiv preprint arXiv:2406.09414, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2.arXiv preprint arXiv:2406.09414, 2024. 3 10

    2024 arXiv

  46. [54]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiao- han Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024. 1, 3, 7

    2024 arXiv

  47. [55]

    Stedge: Self-training edge detection with multilayer teaching and regularization.IEEE Transactions on Neural Networks and Learning Systems, 2023

    Yunfan Ye, Renjiao Yi, Zhiping Cai, and Kai Xu. Stedge: Self-training edge detection with multilayer teaching and regularization.IEEE Transactions on Neural Networks and Learning Systems, 2023. 3

    2023

  48. [56]

    Diffusionedge: Diffusion probabilistic model for crisp edge detection

    Yunfan Ye, Kai Xu, Yuhang Huang, Renjiao Yi, and Zhiping Cai. Diffusionedge: Diffusion probabilistic model for crisp edge detection. InProceedings of the AAAI Conference on Artificial Intelligence, pages 6675–6683, 2024. 3

    2024

  49. [57]

    Zixiao Yu, Lele Fu, Yongyong Chen, Zhiling Cai, and Guo- qing Chao. Hyper-laplacian regularized concept factoriza- tion in low-rank tensor space for multi-view clustering.IEEE Transactions on Emerging Topics in Computational Intelli- gence, pages 1–15, 2024. 2

    2024

  50. [58]

    Identity- preserving text-to-video generation by frequency decompo- sition

    Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyuan Ge, Yu- jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity- preserving text-to-video generation by frequency decompo- sition. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2025. 2

    2025

  51. [59]

    Multi- scale video anomaly detection by multi-grained spatio- temporal representation learning

    Menghao Zhang, Jingyu Wang, Qi Qi, Haifeng Sun, Zirui Zhuang, Pengfei Ren, Ruilong Ma, and Jianxin Liao. Multi- scale video anomaly detection by multi-grained spatio- temporal representation learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Reco...

    2024

  52. [60]

    Pose-guided transformer for fine-grained action quality assessment.IEEE Transactions on Circuits and Systems for Video Technology, pages 1–1,

    Yanting Zhang, Xia Li, Wenhao Chai, Cairong Yan, Wen- hai Wang, and Gaoang Wang. Pose-guided transformer for fine-grained action quality assessment.IEEE Transactions on Circuits and Systems for Video Technology, pages 1–1,

  53. [62]

    (11) Final Score:The warping errorE warp is calculated as the average of the pixel-wise differences over all consecutive frame pairs: Ewarp = 1 T−1 T−1X t=1 Et, (12) whereTdenotes the total number of frames.For specific ex- amples, please refer to Fig. 7. A.2. Appearance anoma...

  54. [2025]

    1 11 HumanSAM: Classifying Human-centric Forgery Videos in Human Spatial, Appearance, and Motion Anomaly Supplementary Material In this supplementary material, we offer further de- tails on HumanSAM. Sec. A delves into the calculation of the three anomaly scoring mechanisms. S...

Pith tools

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