Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Deepfake Detection with Spatio-Temporal Consistency and Attention

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

Pith's one-line read Deepfake videos are best detected by their local texture errors and inter-frame motion inconsistencies, and this paper's dual-attention ResNet50 reports 90.91% accuracy on FF++(LQ) and 95.97% cross-dataset on DFDC with lower compute than…

desk verdict A sensible architecture with an unverifiable central performance claim: Table I and Table II contradict each other on the same dataset. read the letter →

arxiv 2502.08216 v1 pith:DY5FZOIB submitted 2025-02-12 cs.CV

classification cs.CV
keywords deepfakedetectionspatio-temporalattentionopticalflowvideoforensicsFaceForensics++DFDCfine-grainedclassificationtextureenhancement
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

Deepfake generators usually process videos frame by frame, so the tell-tale defects they leave are local texture errors inside frames and incoherent motion between frames. This paper proposes a detector built around those two signals: a ResNet50 backbone whose shallow features are sharpened by texture enhancement and spatial attention, plus a temporal stream that computes optical flow between consecutive frames and runs the motion residuals through a vision-transformer distance attention to produce temporal attention maps. The two attention streams are fused and a binary classifier decides real versus fake. On FaceForensics++(LQ) the paper reports 90.91% accuracy and 0.9810 AUC, and on a cross-dataset test from FF++ to DFDC it reports 95.97% accuracy, with fewer parameters, less memory, and lower FLOPs than eight comparison methods. The point of the work is that modeling fine-grained spatio-temporal consistency, rather than global frame content, is a more effective and cheaper way to detect deepfakes.

What carries the argument

The load-bearing mechanism is a dual attention fusion over space and time. In the spatial stream, a Dense-block texture enhancement amplifies shallow features, and a WS-DAN spatial attention map multiplies the shallow feature map so the network concentrates on localized artifacts. In the temporal stream, optical flow between consecutive frames yields motion residuals; the residuals are split into 3×3 patches, embedded, and processed by a vision-transformer distance attention mechanism whose learned 'global fake template' produces temporal attention maps. The spatial and temporal attention maps guide feature maps at different depths and are fused before the classifier, forcing the network to answer both where the manipulation is and when the motion becomes incoherent.

What would settle it

Compute the accuracy implied by Table I's confusion matrix ($(1869+1943)/3886 = 98.1\%$) and compare it with the 90.91% ACC reported in Table II for the same FF++(LQ) evaluation; then run a controlled reproduction of the FF++ to DFDC test with the same train/validation split, preprocessing, and identity-overlap controls used for the baselines. If the two accuracy figures cannot be reconciled, or if the cross-dataset advantage shrinks or reverses under a matched protocol, the central claim would not survive.

Watch

Extended reading notes

Core claim

The paper's central claim is that the local signatures of frame-by-frame face manipulation—texture artifacts within a single frame and flicker/jitter between adjacent frames—carry more discriminative information than global frame features, and that a network designed to attend to exactly those regions can outperform existing detectors. Concretely, the proposed model couples a texture enhancement block (a Dense block) with a WS-DAN-style spatial attention module on the ResNet50 backbone to reweight shallow frame features, while a temporal attention module feeds optical-flow motion residuals into a patch-based vision-transformer distance attention mechanism whose activations form temporal attention maps. These are fused with deeper features and classified as real or fake. The author reports this design beats eight state-of-the-art methods on FF++(LQ) in both accuracy and AUC, and transfers better to DFDC than the compared methods, while being lighter in memory and compute.

Load-bearing premise

The central claim stands on the evaluation protocol being fair and uniform: Table II lists 90.91% ACC while Table I's confusion matrix implies 98.1% on the same FF++(LQ) set, and the 95.97% DFDC result in Table III is reported without the test split, preprocessing, decision rule, or identity-overlap checks, so any unaccounted difference between protocols would break the comparison.

Editorial extensions

If this is right

  • On the paper's reported numbers, the model's 25.56 MB parameter count and 4,120 MFLOPS make the spatio-temporal attention design a viable candidate for deployment settings where 3D-CNN or recurrent baselines are too heavy.
  • The optical-flow temporal stream implies that deepfake detectors should treat inter-frame consistency as a first-class signal, not just aggregate frame-level predictions.
  • The reported cross-dataset result of 95.97% from FF++ to DFDC suggests spatio-temporal cues learned on one generator family can transfer to unseen generators and datasets.
  • Fusing temporal attention maps into deeper layers, rather than only at the output, changes where and how the network looks; the paper's attention maps indicate tampered facial regions are marked in both single frames and frame sequences.

Reading between the lines

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

  • Inferred: a direct ablation varying the number of temporal context frames $n$ and the compression level of the input would show how much of the gain comes from the temporal stream versus the spatial attention; the paper's reported numbers are for the fused model as a whole.
  • Inferred: because the temporal signal is optical-flow-based, the method's behavior under low bitrate compression or high frame rate is a natural testable boundary—motion estimation degrades exactly where deepfake flicker is hardest to see.
  • Inferred: the same dual-attention template could transfer to other frame-by-frame synthesis artifacts, for example full-body reenactment or speech-driven lip sync, where the generative model is applied independently to frames and therefore leaves the same temporal incoherence.
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 a deepfake video detection method that combines a ResNet50 backbone with a texture enhancement block, a spatial attention module (based on WS-DAN), and a temporal attention module that uses optical flow and a ViT-based distance attention mechanism. The authors claim state-of-the-art accuracy and AUC on FaceForensics++ (LQ) and DFDC, along with memory and computational advantages over existing methods. The empirical support consists of a confusion matrix (Table I), a comparison table on FF++(LQ) (Table II), a cross-dataset table (Table III), visualizations of attention maps, and an ablation-style discussion.

Significance. If the reported results were reproducible, the method would be a useful contribution: it addresses a fine-grained, localized view of deepfake artifacts and reports lower memory (109.69 MB) and MFLOPS (4120) than several baselines. The cross-dataset result on DFDC, if valid, would be notable. However, the significance is substantially undermined by internal inconsistencies in the reported numbers and by the absence of a clearly defined evaluation protocol, which prevents verification of any of the central claims.

major comments (4)
  1. [Section IV-D, Tables I and II] The confusion matrix in Table I on the FF++(LQ) validation set implies an accuracy of (1869+1943)/3886 = 98.1%, while Table II reports ACC 90.91% for the same model on FF++(LQ). The paper does not reconcile this discrepancy or explain whether these numbers correspond to different splits or different evaluation conditions. This internal inconsistency directly undermines the headline performance claim.
  2. [Section IV-C, Table III] Section IV-C describes only an 80:20 train/validation split and never defines a test split. Table III nevertheless reports 'FF++(LQ) Test Results' and 'DFDC Test Results' without specifying how the test sets were constructed, whether decisions are made at frame level or video level, how faces were detected and cropped at test time, or how the DFDC subset was selected. Without this information, the reported results are not reproducible and the comparison to baselines is not meaningful.
  3. [Table III, DFDC column] The cross-dataset result of 95.97% ACC when training only on FF++(LQ) and testing on DFDC is both unexplained and surprising, as it exceeds the within-dataset accuracy on FF++(LQ) (90.91%). The paper provides no details on the DFDC test subset, its real/fake balance, resolution, or frame selection, nor any check for identity overlap with FF++ training actors. This raises the possibility of a different evaluation protocol or data leakage, and the claim cannot be assessed without a full description.
  4. [Tables II and III] The baselines are not evaluated under a common protocol: the reported numbers are taken from different publications, and even the same method appears with different scores in the two tables (e.g., Xception is 81.00% ACC in Table II but 90.03% in Table III). The claim of 'significant performance over the state-of-the-art methods' requires all models to be evaluated on the same splits and with the same decision rules; otherwise the comparison is invalid.
minor comments (5)
  1. [Section III-C, Eq. (2)] Equation (2) contains a typo: 'I(x, y, z)' should be 'I(x, y, t)'. In addition, the derivation says 'bring Eq. (1) into Eq. (2)', but Eq. (1) is not used explicitly; the optical flow derivation should be rewritten for clarity.
  2. [Section II] The paragraph describing Wu et al. [43] is duplicated verbatim ('Wu et al. [43] proposed that spatial and temporal features can be used as detection targets...'). One copy should be removed.
  3. [References] References [21] and [22] are the same paper (an arXiv preprint and a CVPRW version of 'Exposing Deepfake Videos by Detecting Face Warping Artifacts'). They should be merged or clearly cross-referenced.
  4. [Table II] The memory and MFLOPS columns have missing entries for most baselines, which weakens the claim of memory and computational advantages. The authors should either provide these values or explicitly state why they are unavailable.
  5. [Section III-C] The term 'global fake template' is introduced without any formal definition of how it is learned or what it represents in the latent space. A precise description or a reference is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the paper's claims are empirical benchmark results produced by a trained network, not quantities defined in terms of their own outputs.

full rationale

The paper makes no derivation claim that reduces to its inputs. The detector is a ResNet50 backbone augmented with a Dense-block texture enhancer, WS-DAN-style spatial attention, and a ViT-style temporal distance attention over optical-flow residuals; all components are standard, externally defined architectures. Reported ACC/AUC values in Tables II and III are test-set outputs of a classifier trained on FF++(LQ), and the confusion matrix in Table I is also an empirical validation-set result, so no parameter is fitted to the metric it is said to predict. The paper contains no self-citations, uniqueness theorem, or ansatz imported from the authors' prior work. The internal inconsistency between Table I's implied 98.1% validation accuracy and Table II's 90.91% accuracy, and the under-specified DFDC test protocol, are evaluation and reporting concerns that bear on correctness and reproducibility, not on circularity. In the absence of any step where an output is defined in terms of the claimed result or where a fitted quantity is renamed as a prediction, the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central empirical claim depends on several unstated protocol choices and hand-picked hyperparameters, plus domain assumptions about how deepfakes manifest. The only invented entity is the global fake template, a learned internal vector without independent evidence.

free parameters (4)
  • n (number of future frames for temporal attention) = not reported
    The paper states the temporal attention module needs the next n frames but never gives the value or a sensitivity study.
  • face crop sizes = 398x398 and 224x224
    Chosen without ablation; the input resolution to backbone and attention modules is a hand-selected hyperparameter.
  • training iterations = 80
    Iteration count is fixed at 80 with early stopping; no analysis of convergence or sensitivity is provided.
  • ViT sequence length = 9 (3x3)
    Patch grid size is set to 3x3 without experiments on other partitions.
assumptions (5)
  • domain assumption Deepfake videos are generated frame by frame, creating temporal inconsistencies (flicker and jitter) and localized spatial artifacts.
    This premise motivates the entire spatio-temporal attention design in Section III.
  • standard math Brightness constancy and first-order Taylor expansion are valid for computing optical flow between adjacent frames.
    Used in Eqs. (1)-(4) to derive the optical flow constraint; standard in computer vision.
  • domain assumption Dlib face detection and alignment produce accurate, consistent face crops over the videos in FF++ and DFDC.
    The preprocessing pipeline in Section IV-C depends on this; failures would corrupt both training and test inputs.
  • domain assumption The train/test splits of FF++(LQ) and DFDC are used correctly and the comparison baselines' reported numbers come from comparable protocols.
    The empirical claims in Tables II and III rely on this; the paper provides no leakage checks or protocol harmonization.
  • domain assumption MSE loss is an appropriate objective for binary classification of real versus fake.
    Adopted in Section IV-C without justification; for a two-class problem cross-entropy is the standard choice.
invented entities (1)
  • Global fake template
    purpose: A learned latent vector used in the temporal attention module to produce attention weights from patch embeddings, intended to represent a generic forgery pattern.
    The template is internal to the model and is not validated by any external measurement; its only support is the final classification accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deepfake Detection with Spatio-Temporal Consistency and Attention." pith.science (2026). https://pith.science/paper/DY5FZOIB

@misc{pith2026250208216,
  author       = {Pith},
  title        = {Pith review of: Deepfake Detection with Spatio-Temporal Consistency and Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DY5FZOIB}},
  note         = {Machine review of arXiv:2502.08216}
}
read the original abstract

Deepfake videos are causing growing concerns among communities due to their ever-increasing realism. Naturally, automated detection of forged Deepfake videos is attracting a proportional amount of interest of researchers. Current methods for detecting forged videos mainly rely on global frame features and under-utilize the spatio-temporal inconsistencies found in the manipulated videos. Moreover, they fail to attend to manipulation-specific subtle and well-localized pattern variations along both spatial and temporal dimensions. Addressing these gaps, we propose a neural Deepfake detector that focuses on the localized manipulative signatures of the forged videos at individual frame level as well as frame sequence level. Using a ResNet backbone, it strengthens the shallow frame-level feature learning with a spatial attention mechanism. The spatial stream of the model is further helped by fusing texture enhanced shallow features with the deeper features. Simultaneously, the model processes frame sequences with a distance attention mechanism that further allows fusion of temporal attention maps with the learned features at the deeper layers. The overall model is trained to detect forged content as a classifier. We evaluate our method on two popular large data sets and achieve significant performance over the state-of-the-art methods.Moreover, our technique also provides memory and computational advantages over the competitive techniques.

Figures

Figures reproduced from arXiv: 2502.08216 by the authors.

Figure 1
Figure 1. The proposed method leverages three major components. (a) Attention [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of our method consists of three important components. A texture enhancement block module for enchancing the texture features. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Temporal attention incorporation in the backbone. A single frame and [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Optical flow [10] of the motion residual describes the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: The employed long-distance attention mechanism. This mechanism divides the input into patches and treats these patches as a sequence(3×3). The [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: Attention maps generated by consecutive frames. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 6
Figure 6. Figure 6: Attention maps generated by single frames. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 34 canonical work pages

  1. [1]

    Mesonet: a compact facial video forgery detection network

    Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1–7. IEEE, 2018

  2. [2]

    A deep learning approach to universal image manipulation detection using a new convolutional layer

    Belhassen Bayar and Matthew C Stamm. A deep learning approach to universal image manipulation detection using a new convolutional layer. In Proceedings of the 4th ACM workshop on information hiding and multimedia security , pages 5–10, 2016

  3. [3]

    Deepfakes: what are they and why would i make one? URL https://www

    BBC Bitesize. Deepfakes: what are they and why would i make one? URL https://www. bbc. co. uk/bitesize/articles/zfkwcqt , 2019

  4. [4]

    Jumping nlp curves: A review of natural language processing research [review article]

    Erik Cambria and Bebo White. Jumping nlp curves: A review of natural language processing research [review article]. IEEE Computational Intelligence Magazine, 9(2):48–57, 2014

  5. [5]

    Xception: Deep learning with depthwise separable convolutions

    Franc ¸ois Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1251–1258, 2017

  6. [6]

    The deepfake detection challenge (dfdc) dataset

    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

  7. [7]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2016

  8. [8]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 770–778, 2016

Show all 45 references
  1. [9]

    Long short-term memory

    Sepp Hochreiter and J ¨urgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997

  2. [10]

    Determining optical flow

    Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981

  3. [11]

    Squeeze-and-excitation networks

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132–7141, 2018

  4. [12]

    See better before looking closer: Weakly supervised data augmentation network for fine- grained visual classification

    Tao Hu, Honggang Qi, Qingming Huang, and Yan Lu. See better before looking closer: Weakly supervised data augmentation network for fine- grained visual classification. arXiv preprint arXiv:1901.09891 , 2019

  5. [13]

    Weinberger

    Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017

  6. [14]

    All tokens matter: Token labeling for training better vision transformers

    Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. Advances in Neural Information Processing Systems, 34:18590–18602, 2021

  7. [15]

    Transfiguring portraits

    Ira Kemelmacher-Shlizerman. Transfiguring portraits. ACM TOG , 35(4):1–8, 2016

  8. [16]

    Dlib-ml: A machine learning toolkit

    Davis E King. Dlib-ml: A machine learning toolkit. The Journal of Machine Learning Research , 10:1755–1758, 2009

  9. [17]

    Head2head: Video-based neural head synthesis

    Mohammad Rami Koujan, Michail Christos Doukas, Anastasios Rous- sos, and Stefanos Zafeiriou. Head2head: Video-based neural head synthesis. In FG 2020, pages 16–23. IEEE, 2020

  10. [18]

    Face x-ray for more general face forgery detection

    Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5001–5010, 2020

  11. [19]

    Exposing deepfake videos by tracking eye movements

    Meng Li, Beibei Liu, Yongjian Hu, and Yufei Wang. Exposing deepfake videos by tracking eye movements. In 2020 25th International Conference on Pattern Recognition (ICPR) , pages 5184–5189. IEEE, 2021

  12. [20]

    In ictu oculi: Exposing ai created fake videos by detecting eye blinking

    Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International workshop on information forensics and security (WIFS) , pages 1–7. IEEE, 2018

  13. [21]

    Exposing deepfake videos by detecting face warping artifacts

    Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. arXiv preprint arXiv:1811.00656 , 2018

  14. [22]

    Exposing deepfake videos by detecting face warping artifacts

    Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019

  15. [23]

    Exposingaicreated fakev- ideosbydetectingeyeblinking

    CHANG M LIY and LYUS InIctuOculi. Exposingaicreated fakev- ideosbydetectingeyeblinking. In 2018IEEEInterG national Workshop on Information Forensics and Security (WIFS). IEEE , 2018

  16. [24]

    Detection of deepfake videos using long distance attention

    Wei Lu, Lingyi Liu, Junwei Luo, Xianfeng Zhao, Yicong Zhou, and Jiwu Huang. Detection of deepfake videos using long distance attention. arXiv preprint arXiv:2106.12832 , 2021

  17. [25]

    Two-branch recurrent network for isolating deepfakes in videos

    Iacopo Masi, Aditya Killekar, Royston Marian Mascarenhas, Shenoy Pratik Gurudatt, and Wael AbdAlmageed. Two-branch recurrent network for isolating deepfakes in videos. In European conference on computer vision , pages 667–684. Springer, 2020

  18. [26]

    Exploiting visual artifacts to expose deepfakes and face manipulations

    Falko Matern, Christian Riess, and Marc Stamminger. Exploiting visual artifacts to expose deepfakes and face manipulations. In 2019 IEEE Winter Applications of Computer Vision Workshops (WACVW) , pages 83–92. IEEE, 2019

  19. [27]

    Rsgan: face swapping and editing using face and hair representation in latent spaces

    Ryota Natsume, Tatsuya Yatagawa, and Shigeo Morishima. Rsgan: face swapping and editing using face and hair representation in latent spaces. arXiv preprint arXiv:1804.03447 , 2018

  20. [28]

    Multi-task learning for detecting and segmenting manipulated facial images and videos

    Huy H Nguyen, Fuming Fang, Junichi Yamagishi, and Isao Echizen. Multi-task learning for detecting and segmenting manipulated facial images and videos. In 2019 IEEE 10th International Conference on Biometrics Theory, Applications and Systems (BTAS) , pages 1–8. IEEE, 2019

  21. [29]

    Capsule-forensics: Using capsule networks to detect forged images and videos

    Huy H Nguyen, Junichi Yamagishi, and Isao Echizen. Capsule-forensics: Using capsule networks to detect forged images and videos. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 2307–2311. IEEE, 2019

  22. [30]

    Fsgan: Subject agnostic face swapping and reenactment

    Yuval Nirkin, Yosi Keller, and Tal Hassner. Fsgan: Subject agnostic face swapping and reenactment. In ICCV, pages 7184–7193, 2019

  23. [31]

    An introduction to convolutional neural networks

    Keiron O’Shea and Ryan Nash. An introduction to convolutional neural networks. arXiv preprint arXiv:1511.08458 , 2015

  24. [32]

    Ganimation: Anatomically-aware facial animation from a single image

    Albert Pumarola, Antonio Agudo, Aleix M Martinez, Alberto Sanfeliu, and Francesc Moreno-Noguer. Ganimation: Anatomically-aware facial animation from a single image. In ECCV, 2018

  25. [33]

    Porn producers offer to help hollywood take down deepfake videos

    Janko Roettgers. Porn producers offer to help hollywood take down deepfake videos. Variety. February, 21, 2018

  26. [34]

    Faceforensics++: Learning to detect manipulated facial images

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

  27. [35]

    Efficient residual dense block search for image super-resolution

    Dehua Song, Chang Xu, Xu Jia, Yiyi Chen, Chunjing Xu, and Yunhe Wang. Efficient residual dense block search for image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence , vol- ume 34, pages 12007–12014, 2020

  28. [36]

    Synthesizing obama: learning lip sync from audio

    Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher- Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1–13, 2017

  29. [37]

    Face2face: Real-time face capture and reenact- ment of rgb videos

    Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenact- ment of rgb videos. In Proceedings on CVPR, pages 2387–2395, 2016

  30. [38]

    Deepfakes and beyond: A survey of face manipulation and fake detection

    Ruben Tolosana, Ruben Vera-Rodriguez, Julian Fierrez, Aythami Morales, and Javier Ortega-Garcia. Deepfakes and beyond: A survey of face manipulation and fake detection. Information Fusion , 64:131– 148, 2020

  31. [39]

    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. In Proceedings of the IEEE international conference on computer vision, pages 4489–4497, 2015

  32. [40]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

  33. [41]

    The emergence of deepfake technology: A review

    Mika Westerlund. The emergence of deepfake technology: A review. TIM Review, 9(11), 2019

  34. [42]

    Reenactgan: Learning to reenact faces via boundary transfer

    Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. Reenactgan: Learning to reenact faces via boundary transfer. In ECCV, pages 603–619, 2018

  35. [43]

    Sstnet: Detecting manipulated faces through spatial, steganalysis and temporal features

    Xi Wu, Zhen Xie, YuTao Gao, and Yu Xiao. Sstnet: Detecting manipulated faces through spatial, steganalysis and temporal features. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 2952–2956. IEEE, 2020

  36. [44]

    Exposing deep fakes using inconsistent head poses

    Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261–8265. IEEE, 2019

  37. [45]

    Multi-attentional deepfake detection

    Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.