Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content

T0 review · 5 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read UNITE claims that one transformer trained on face swaps plus game footage can detect face manipulations, background edits, and fully AI-generated videos, even with no face in the frame.

desk verdict Genuinely useful attention-diversity loss and a clever GTA-V training trick, but the universal-detector claim outruns the evidence—still worth serious peer review. read the letter →

arxiv 2412.12278 v2 pith:J3YPABCA submitted 2024-12-16 cs.CV

Rohit Kundu , Hao Xiong , Vishal Mohanty , Athula Balachandran , Amit K. Roy-Chowdhury This is my paper · ORCID
classification cs.CV
keywords deepfakedetectionsyntheticvideotext-to-videoimage-to-videoattentiondiversitylosscross-domaingeneralizationSigLIPforensics
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 tries to build a single video-forensics model that works across the full spectrum of synthetic manipulation: classic face swaps and lip-syncs, background-only edits, and fully AI-generated clips from text-to-video (T2V) and image-to-video (I2V) models. The motivating problem is that today's detectors are face-centric—they often cannot even run without a detected face—while modern generators can alter scenery or produce entire videos with no humans at all. UNITE processes whole frames through a frozen SigLIP foundation encoder feeding a small trainable transformer, and it is trained on FaceForensics++ plus the GTA-V game dataset, the latter standing in for AI-generated video despite being rendered rather than generated. The key move is an attention-diversity loss that stops the transformer's heads from all fixating on faces, forcing the model to spread its scrutiny across foreground and background. If the claim holds, a single detector can replace the current patchwork of separate DeepFake and synthetic-video detectors.

What carries the argument

The mechanism has three parts. First, a frozen SigLIP-So400M image encoder (a shape-optimized vision transformer contrastively pretrained with a sigmoid loss) converts each resized 384×384 frame into 729 tokens of dimension 1152; the tokens for 64 sampled frames are concatenated in temporal order. Second, a small learnable transformer (4 encoder blocks, 12 attention heads, sine-cosine positional encoding) ingests this token sequence and learns temporal inconsistencies; attention maps from the first encoder block are also used as interpretability and as input to the loss. Third, the attention-diversity (AD) loss, combining a within-class term that pulls attention-pooled features toward per-class feature centers and a between-class term that pushes centers apart, prevents all heads from collapsing onto the face region. The training mix is FaceForensics++ (face manipulations) plus SAIL-VOS-3D (GTA-V game footage), which is fully synthetic but not AI-generated; the paper's argument is that this non-AI synthetic corpus teaches the model the general artifact signature of generated content in SigLIP feature space.

What would settle it

Train UNITE exactly as described (FF++ plus GTA-V, CE+AD loss), then evaluate it on videos from a T2V/I2V generator family not present in DeMamba—for example, Kling, Veo, or a newer Sora checkpoint. If accuracy on such unseen-generator videos falls to near chance while a detector trained on DeMamba data stays high, then the GTA-V proxy claim fails. A complementary check: compute SigLIP-embedding distances between GTA-V frames and held-out T2V/I2V frames and show whether they overlap enough to justify the proxy.

Watch

Extended reading notes

Core claim

The paper's central claim is that UNITE—a full-frame, transformer-based detector—can identify partially manipulated (foreground or background) and fully synthetic videos in cross-domain settings with a single model. On face-manipulated benchmarks (FF++, CelebDF, DeeperForensics, UADFV, HifiFace, DeepfakeTIMIT) it matches or exceeds specialized face-cropping detectors. On the DeMamba benchmark, which contains videos from ten T2V/I2V generators including Sora, UNITE outperforms the current synthetic-video detectors' reported precision and recall averages even though, unlike them, it was not trained on any DeMamba data. The main supporting experiment is the contrast between training with FF++ alone and training with FF++ plus GTA-V: adding the game footage lifts accuracy on background-manipulated AVID videos from 41.67% to 100%, on DeMamba from 61.47% to 87.12%, and on CelebDF from 72.61% to 95.11%. The paper also reports that a fine-grained three-class version can separate real, partially manipulated, and fully AI-generated videos.

Load-bearing premise

The load-bearing premise is that GTA-V game footage, though not produced by AI, leaves artifacts in SigLIP feature space that stand in for the artifacts of real T2V/I2V generators; if game-rendering artifacts are too different from AI-generation artifacts, the reported cross-domain gains would not transfer to unseen generators.

Editorial extensions

If this is right

  • One trained model replaces separate DeepFake and T2V/I2V detectors, since UNITE handles both categories at once.
  • Detection no longer requires a visible face, so background inpainting, non-human scenes, and animal or object videos become checkable.
  • Adding non-AI synthetic game data to training improves not only synthetic-video detection but also cross-dataset face-manipulation performance, e.g., +22.5 accuracy on CelebDF.
  • A three-way variant can triage videos as real, partially manipulated, or fully AI-generated, offering explainability beyond a binary fake flag.
  • Attention heatmaps become broader and less face-only when AD loss is used, suggesting the model is actually looking at manipulation-relevant regions.

Reading between the lines

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

  • The proxy strategy implies a broader hypothesis: any sufficiently diverse synthetic-but-not-AI video corpus (other game engines, rendered 3D scenes, CGI films) might substitute for GTA-V; a direct test would retrain UNITE on a different rendering source and check whether DeMamba and AVID gains persist.
  • Because AD loss is what produces the spatial attention spread, the same loss may make attention maps usable as weak localization for manipulated regions—something the paper visualizes but does not formally evaluate.
  • A risk the paper does not discuss: a detector trained on GTA-V may misclassify real video-game footage or CGI in the wild as fake; this is a testable failure mode.
  • The frames-versus-performance result (accuracy rises from roughly 50% with 1 frame to 87% with 64 on DeMamba) suggests temporal inconsistency is a major cue, so a frame-level static detector would likely underperform on the same benchmarks—an implicit comparison the paper does not run.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes UNITE, a transformer-based detector aimed at face-manipulated, background-manipulated, and fully synthetic T2V/I2V videos. Video frames are encoded with the frozen SigLIP-So400M foundation model, and a four-block transformer with multi-head attention is trained with a cross-entropy loss plus an attention-diversity (AD) loss intended to spread spatial attention beyond faces. Training uses FaceForensics++ and the SAIL-VOS-3D dataset (GTA-V game footage, which the paper explicitly notes is not AI-generated). Cross-dataset evaluations cover several face-manipulation benchmarks, the AVID background-inpainting dataset, the DeMamba T2V/I2V benchmark, and a 10-video New York Times quiz. The paper reports strong cross-domain results and claims that a single trained model eliminates the need for separate DeepFake and T2V/I2V detectors.

Significance. If the claims hold, the paper offers a useful step toward a single full-frame detector for diverse video manipulations, and the recipe of frozen foundation features plus non-AI game footage as a synthetic proxy is an interesting and cost-effective idea. The breadth of the evaluation across face, background, and fully synthetic data is a strength, as are the ablations on frame count, transformer depth, and loss components. However, the two most distinctive claims—background manipulation detection and fully synthetic detection—rest on very small evaluation sets and on a single proxy-training transfer benchmark, and the AD-loss formulation as written is not reproducible. The central 'universal' claim is therefore not yet established; the paper needs additional evidence and substantial clarification before it can be accepted as a journal-level contribution.

major comments (5)
  1. [Section 3.4, Eqs. (3)-(5)] The feature-center formulation is not well-defined. C is declared as R^{nh×nf} and Eq. (3) updates it from the batch-averaged pooled feature with no class conditioning, yet Eq. (4) needs a per-class center for each sample and Eq. (5) forms pairwise distances between centers of different classes. As written, Eq. (3) yields a single global center (or one per head), not the per-class anchors required by Lwithin and Lbetween. Also, δwithin=[0.01,-2] for binary classification gives one class a negative allowed distance. Please correct the notation and update rule so the loss is computable, and clarify how class-conditioned centers are maintained.
  2. [Section 4.4 and Table 3] The paper's central claim that non-AI GTA-V game data transfers to AI-generated video is not established. Section 4.4 states that on DeMamba, CE-only performance is similar regardless of whether GTA-V is included in training; the accuracy boost appears only when the AD loss is added. This suggests the improvement may be an interaction between the regularizer and this particular benchmark rather than evidence of a transferable synthetic-video artifact. Moreover, Table 3 contains only a fixed set of DeMamba generators; there is no held-out T2V/I2V generator. Please add an unseen-generator evaluation or feature-space evidence for the GTA-V-to-T2V/I2V transfer, or weaken the universal claim accordingly.
  3. [Table 1 (AVID and NYTimes rows)] The two most distinctive evaluations are based on very small test sets, but the sample sizes are not reported. The AVID row jumps from 41.67% to 100% accuracy; the percentages imply a test set on the order of a dozen videos. The NYTimes evaluation uses only 10 videos, of which UNITE gets 8 right. These numbers cannot support a reliable cross-domain claim, and the 100% figure is particularly fragile. Please report exact numbers of test videos, per-video predictions, and confidence intervals, and avoid point-estimate claims on such small samples.
  4. [Section 4.2, Table 2] The SOTA comparison on face-manipulated data is not matched. UNITE is trained on FF++ plus GTA-V, while most baselines in Table 2 are trained on FF++ only (or on their own protocols). The large gains on CelebDF and DeeperForensics could be due to the additional training data rather than the architecture or the AD loss. Please state the training data for each baseline and provide a matched comparison where all methods are trained on the same data.
  5. [Section 4.2, Training Details] No validation protocol is described. The hyperparameters of the AD loss (δwithin, δbetween, λ1, λ2, η), the transformer depth, and the number of frames are fixed, but the paper does not say whether they were chosen on a held-out validation split or on the same cross-domain sets used for the final tables. Without this information, the reported cross-dataset gains are at risk of selection bias. Please specify the validation procedure and include sensitivity analysis for all AD-loss hyperparameters.
minor comments (5)
  1. [Section 3.3.2, Eq. (1)] The positional encoding is defined only over the frame index j and does not distinguish the ts=729 spatial tokens within a frame. Please clarify how each token receives its positional information; as written, all tokens of a frame share the same encoding, which weakens the intended temporal modeling.
  2. [Section 4.3, Table 4] The fine-grained classes are inconsistent with the paper's own definition: GTA-V is repeatedly described as synthetic but not AI-generated, yet it is placed in the 'fully synthetic' class alongside DeMamba. Also, the 0.00% GTA-V accuracy in the FF++-only row is surprising given the 60.16% binary result in Table 1; please explain this discrepancy.
  3. [Section 4.2 and Section 4.4] There are several typos, including 'by by handling' in Section 4.2 and 'sythetic' in Section 4.4. Please proofread the manuscript.
  4. [General] The paper does not state whether code or model weights will be released. For a detection method whose main novelty is a training recipe and a loss function, code release would substantially improve reproducibility.
  5. [Table 3] Please state the threshold at which UNITE's precision/recall values in Table 3 are computed and confirm that the same 0.5 threshold is used as in the other tables; the DeMamba baselines may use different decision rules.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: UNITE's claims are empirical cross-domain results with training and evaluation sets disjoint, and no load-bearing self-citation or definitional reduction is present.

full rationale

The paper contains no derivation chain whose outputs are equivalent to its inputs by construction. UNITE is an empirical detector: frozen SigLIP features are fed to a learned transformer, trained with cross-entropy plus an attention-diversity loss on FF++ and GTA-V, and then evaluated on external benchmarks including CelebDF, DeeperForensics, AVID, the NYTimes quiz, and the DeMamba validation split. None of the method components is defined in terms of the target evaluation metrics, and no fitted parameter is relabeled as a prediction. The use of GTA-V as a proxy for AI-generated content is explicitly acknowledged as not AI-generated in both the introduction and conclusion; this raises an external-validity concern about whether game-rendering artifacts transfer to T2V/I2V artifacts, not a circularity concern, since the transfer is an empirical hypothesis tested on disjoint data rather than an analytic identity. The comparison against DeMamba detectors relies on results reported by Chen et al., which is an independent external benchmark, and UNITE's advantage there is not forced by construction. The only self-citations (e.g., Mazaheri et al. in related work) are descriptive and not load-bearing for the paper's central contribution. Hyperparameter choices such as AD-loss margins, transformer depth, and loss weights are not shown to have been selected on the evaluation sets, and speculative tuning risk without quoted evidence does not meet the standard for a circularity finding. Overall, the central claims are self-contained empirical results against external benchmarks, so the appropriate circularity score is 0.

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

The central claim rests on the transferability of synthetic game footage to AI-generated video, on segment-level label inheritance, and on the assumption that first-block attention maps are the right place to enforce spatial diversity. No new physical or conceptual entities are introduced.

free parameters (4)
  • delta_within = [0.01, -2] binary; [0.01, -2, 1] fine-grained
    Within-class margin in Eq 4; hand-set, sensitivity deferred to supplementary.
  • delta_between = 0.5
    Between-class margin in Eq 5; hand-set.
  • lambda1/lambda2 loss weights = 0.5/0.5
    Weights on CE and AD losses in Eq 6; hand-set.
  • feature center learning rate eta = 0.05
    Update rate for attention-pooled feature centers in Eq 3; hand-set.
assumptions (4)
  • domain assumption GTA-V synthetic game footage, though not AI-generated, is a sufficient proxy for AI-generated video artifacts.
    Sec 4.1 and Sec 5 use SAIL-VOS-3D/GTA-V as training data to simulate AI-generated media; if this transfer fails, the reported DeMamba gains lose their explanation.
  • domain assumption Every 64-frame segment of a video inherits the full-video label.
    Sec 3.1 assigns segment labels from source video labels, which assumes manipulations span entire videos.
  • ad hoc to paper Attention maps from the first encoder block reflect the spatial evidence used for classification.
    Sec 3.4 uses first-block attention maps to compute the AD loss; if later blocks make the decision, the loss may not act on the decisive spatial regions.
  • ad hoc to paper The feature-center update rule (Eq 3) with zero initialization converges to stable per-class anchors.
    No convergence analysis is provided; the method assumes the EMA-like update keeps centers meaningful during training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content." pith.science (2026). https://pith.science/paper/J3YPABCA

@misc{pith2026241212278,
  author       = {Pith},
  title        = {Pith review of: Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J3YPABCA}},
  note         = {Machine review of arXiv:2412.12278}
}
read the original abstract

Existing DeepFake detection techniques primarily focus on facial manipulations, such as face-swapping or lip-syncing. However, advancements in text-to-video (T2V) and image-to-video (I2V) generative models now allow fully AI-generated synthetic content and seamless background alterations, challenging face-centric detection methods and demanding more versatile approaches. To address this, we introduce the \underline{U}niversal \underline{N}etwork for \underline{I}dentifying \underline{T}ampered and synth\underline{E}tic videos (\texttt{UNITE}) model, which, unlike traditional detectors, captures full-frame manipulations. \texttt{UNITE} extends detection capabilities to scenarios without faces, non-human subjects, and complex background modifications. It leverages a transformer-based architecture that processes domain-agnostic features extracted from videos via the SigLIP-So400M foundation model. Given limited datasets encompassing both facial/background alterations and T2V/I2V content, we integrate task-irrelevant data alongside standard DeepFake datasets in training. We further mitigate the model's tendency to over-focus on faces by incorporating an attention-diversity (AD) loss, which promotes diverse spatial attention across video frames. Combining AD loss with cross-entropy improves detection performance across varied contexts. Comparative evaluations demonstrate that \texttt{UNITE} outperforms state-of-the-art detectors on datasets (in cross-data settings) featuring face/background manipulations and fully synthetic T2V/I2V videos, showcasing its adaptability and generalizable detection capabilities.

Figures

Figures reproduced from arXiv: 2412.12278 by the authors.

Figure 1
Figure 1. Problem Overview: Existing DeepFake detection meth￾ods primarily focus on identifying face-manipulated videos, most of which cannot perform inference unless there is a face detected in the video. However, with advancements like seamless back￾ground modifications (e.g., AVID [61]) and hyper-realistic con￾tent from games like GTA-V [24] and T2V/I2V models [9], a more comprehensive approach is needed. A model trained w… view at source ↗
Figure 2
Figure 2. UNITE architecture overview: We extract domain￾agnostic features (ξ) using the SigLIP-So400m foundation model [4] to mitigate domain gaps between DeepFake datasets (Sec. 3.2). These embeddings, combined with positional encodings, are input to a transformer with multi-head attention and MLP layers (Sec. 3.3), culminating in a classifier for final predictions. AD-loss (Sec. 3.4) encourages the model attention to span … view at source ↗
Figure 3
Figure 3. Comparison of attention heatmaps when UNITE is trained with only CE loss (second column) versus CE and AD￾losses (third column). The CE-loss heatmap predominantly fo￾cuses on the face region, whereas the CE+AD loss heatmap demonstrates more distributed attention across the entire frame, as indicated by a broader bluish tone in the heatmaps. Both sam￾ples are “fake” videos– (a) is taken from Celeb-DF [30] and (d) is … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Ablation Results to show the effect of changing the loss functions used to train the UNITE model. The combination of the cross-entropy (CE) and Attention Diversity (AD) losses always performs better, and the contribution from the AD-loss component increases significant…
Figure 5
Figure 5. Figure 5: No. of frames vs. Performance: Performance analysis of UNITE based on the number of frames sampled per video seg￾ment. The results illustrate that as the number of frames increases from 1 to 64 (context window), the detection accuracy improves, showcasing UNITE’s abili…
Figure 6
Figure 6. Figure 6: Transformer Depth Evaluation: Performance compar￾ison of UNITE with varying the number of encoder blocks (depth). UNITE performs optimally in cross-domain settings when the depth is 4. Greater depths overfit to the training domains (FF++ [43] and GTA-V [24]), while a d…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. DAVID-XR1: Detecting AI-Generated Videos with Explainable Reasoning

    cs.CV 2025-06 conditional novelty 6.0 of 10

    A video-language model fine-tuned on a new defect-annotated dataset detects AI-generated videos from unseen generators with 76.7% accuracy and gives written explanations, though the test set is small and the dataset i...

Reference graph

Works this paper leans on

64 extracted references · 43 canonical work pages · cited by 1 Pith paper

  1. [1]

    https://blackforestlabs.ai/, 2024

    Flux 1.1. https://blackforestlabs.ai/, 2024. 1

    work page 2024

  2. [2]

    https://www.morphstudio.com/ ,

    Morph studio. https://www.morphstudio.com/ ,

    work page

  3. [3]

    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 informa- tion forensics and security (WIFS) , pages 1–7. IEEE, 2018. 6

    work page 2018

  4. [4]

    Getting vit in shape: Scaling laws for compute-optimal model design

    Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2024. 2, 3

    work page 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]

    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, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video- generation-models-as-world-simulators, 3, 2024. 6, 7

    work page 2024

  7. [7]

    Featuretransfer: Unsu- pervised domain adaptation for cross-domain deepfake de- tection

    Baoying Chen and Shunquan Tan. Featuretransfer: Unsu- pervised domain adaptation for cross-domain deepfake de- tection. Security and Communication Networks , 2021(1): 9942754, 2021. 3

    work page 2021

  8. [8]

    Videocrafter1: Open diffusion models for high-quality video generation

    Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 7

    arXiv

Show all 64 references
  1. [9]

    Demamba: Ai-generated video detection on million-scale genvideo benchmark

    Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. Demamba: Ai-generated video detection on million-scale genvideo benchmark. arXiv preprint arXiv:2405.19707, 2024. 1, 2, 5, 6, 7, 8

    2024 arXiv

  2. [10]

    Can we leave deepfake data behind in training deepfake detector? arXiv preprint arXiv:2408.17052, 2024

    Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector? arXiv preprint arXiv:2408.17052, 2024. 1

    2024 arXiv

  3. [11]

    Exploiting style latent flows for generalizing deepfake video detection

    Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133–1143, 2024. 2, 6

    2024

  4. [12]

    Fakecatcher: Detection of synthetic portrait videos using biological sig- nals

    Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological sig- nals. IEEE transactions on pattern analysis and machine intelligence, 2020. 6

    2020

  5. [13]

    Quality-based artifact modeling for facial deepfake detec- tion in videos

    Sara Concas, Simone Maurizio La Cava, Roberto Casula, Giulia Orr `u, Giovanni Puglisi, and Gian Luca Marcialis. Quality-based artifact modeling for facial deepfake detec- tion in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pag...

    2024

  6. [14]

    Intriguing properties of syn- thetic images: from generative adversarial networks to diffu- sion models

    Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of syn- thetic images: from generative adversarial networks to diffu- sion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , p...

  7. [15]

    On the detection of synthetic images generated by diffusion mod- els

    Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Gio- vanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion mod- els. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p...

    2023

  8. [16]

    Id-reveal: Identity-aware deepfake video detection

    Davide Cozzolino, Andreas R ¨ossler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. Id-reveal: Identity-aware deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision , pages 15108– 15117, 2021. 2

    2021

  9. [17]

    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. 1

    2006 arXiv

  10. [18]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. In ICLR, 2021. 3

    2021

  11. [19]

    Unmasking deepfakes with simple features

    Ricard Durall, Margret Keuper, Franz-Josef Pfreundt, and Janis Keuper. Unmasking deepfakes with simple features. arXiv preprint arXiv:1911.00686, 2019. 2

    1911 arXiv

  12. [20]

    Spatiotemporal incon- sistency learning for deepfake video detection

    Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal incon- sistency learning for deepfake video detection. In Proceed- ings of the 29th ACM international conference on multime- dia, pages 3473–3481, 2021. 7

    2021

  13. [21]

    Controllable guide-space for generalizable face forgery detection

    Ying Guo, Cheng Zhen, and Pengfei Yan. Controllable guide-space for generalizable face forgery detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20818–20827, 2023. 6

    2023

  14. [22]

    Deep fisher faces

    Harald Hanselmann, Shen Yan, and Hermann Ney. Deep fisher faces. In BMVC, 2017. 4

    2017

  15. [23]

    Gaussian error linear units (gelus)

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4

    2016 arXiv

  16. [24]

    Sail-vos 3d: A synthetic dataset and base- lines for object detection and 3d mesh reconstruction from video data

    Yuan-Ting Hu, Jiahong Wang, Raymond A Yeh, and Alexan- der G Schwing. Sail-vos 3d: A synthetic dataset and base- lines for object detection and 3d mesh reconstruction from video data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages ...

    2021

  17. [25]

    Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection

    Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proceedings 9 of the IEEE/CVF conference on computer vision and pattern recognition, pages 2889–2898, 2020. 5, 7

    2020

  18. [26]

    Deepfake-o-meter v2

    Yan Ju, Chengzhe Sun, Shan Jia, Shuwei Hou, Zhaofeng Si, Soumyya Kanti Datta, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-o-meter v2. 0: An open platform for deepfake detection. In 2024 IEEE 7th Inter- national Conference on Multimedia Information Processing ...

    2024

  19. [27]

    Deepfakes: a new threat to face recognition? assessment and detection

    Pavel Korshunov and S ´ebastien Marcel. Deepfakes: a new threat to face recognition? assessment and detection. arXiv preprint arXiv:1812.08685, 2018. 5

    2018 arXiv

  20. [28]

    Faceshifter: Towards high fidelity and occlusion aware face swapping

    Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1

    1912 arXiv

  21. [29]

    Exposing deepfake videos by detecting face warping artif acts

    Y Li. Exposing deepfake videos by detecting face warping artif acts. arXiv preprint arXiv:1811.00656, 2018. 6

    2018 arXiv

  22. [30]

    Celeb-df: A large-scale challenging dataset for deep- fake forensics

    Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deep- fake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 3207– 3216, 2020. 1, 3, 5, 6, 7, 8

    2020

  23. [31]

    Preserving fairness generalization in deepfake detection

    Li Lin, Xinan He, Yan Ju, Xin Wang, Feng Ding, and Shu Hu. Preserving fairness generalization in deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16815–16825, 2024. 6

    2024

  24. [32]

    Ti2net: temporal identity inconsistency network for deepfake detection

    Baoping Liu, Bo Liu, Ming Ding, Tianqing Zhu, and Xin Yu. Ti2net: temporal identity inconsistency network for deepfake detection. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 4691–4700, 2023. 2, 6

    2023

  25. [33]

    Decoupled weight decay regular- ization

    I Loshchilov and F Hutter. Decoupled weight decay regular- ization. ICLR, 2019. 5

    2019

  26. [34]

    Domainforensics: Exposing face forgery across domains via bi-directional adaptation

    Qingxuan Lv, Yuezun Li, Junyu Dong, Sheng Chen, Hui Yu, Huiyu Zhou, and Shu Zhang. Domainforensics: Exposing face forgery across domains via bi-directional adaptation. IEEE Transactions on Information Forensics and Security ,

  27. [35]

    Latte: Latent diffusion transformer for video generation

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

    2024 arXiv

  28. [36]

    Detection and localization of facial expression manipulations

    Ghazal Mazaheri and Amit K Roy-Chowdhury. Detection and localization of facial expression manipulations. In Pro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 1035–1045, 2022. 2

    2022

  29. [37]

    moonvalley.ai

    moonvalley.ai. moonvalley.ai. https://moonvalley. ai/, 2022. 7

    2022

  30. [38]

    Hotshot-XL, 2023

    John Mullan, Duncan Crawbuck, and Aakash Sastry. Hotshot-XL, 2023. 7

    2023

  31. [39]

    Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake de- tection

    Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake de- tection. In Proceedings of the IEEE/CVF Conference on...

    2024

  32. [40]

    Pudd: Towards robust multi-modal prototype-based deepfake de- tection

    Alvaro Lopez Pellicer, Yi Li, and Plamen Angelov. Pudd: Towards robust multi-modal prototype-based deepfake de- tection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3809– 3817, 2024. 2, 6

    2024

  33. [41]

    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. In European conference on com- puter vision, pages 86–103. Springer, 2020. 7

    2020

  34. [42]

    Text driven video generation

    Runway Research. Text driven video generation. https: //research.runwayml.com/gen2, 2023. 1, 7

    2023

  35. [43]

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

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Chris- tian Riess, Justus Thies, and Matthias Nießner. Faceforen- sics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1–11, 2019. 1, 2, 3, 5, 6, 7, 8

    2019

  36. [44]

    Semi-supervised deep domain adaptation for deepfake detec- tion

    Md Shamim Seraj, Ankita Singh, and Shayok Chakraborty. Semi-supervised deep domain adaptation for deepfake detec- tion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1061–1071, 2024. 2

    2024

  37. [45]

    Diffusionfake: Enhancing generalization in deepfake detection via guided stable diffu- sion

    Ke Sun, Shen Chen, Taiping Yao, Hong Liu, Xiaoshuai Sun, Shouhong Ding, and Rongrong Ji. Diffusionfake: Enhancing generalization in deepfake detection via guided stable diffu- sion. NeurIPS, 2024. 1

    2024

  38. [46]

    Improving the efficiency and robustness of deepfakes detec- tion through precise geometric features

    Zekun Sun, Yujie Han, Zeyu Hua, Na Ruan, and Weijia Jia. Improving the efficiency and robustness of deepfakes detec- tion through precise geometric features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3609–3618, 2021. 6

    2021

  39. [47]

    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. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition,...

    2024

  40. [48]

    The New York Times. A.i. can now create lifelike videos. can you tell what’s real? https://www.nytimes.com/ interactive / 2024 / 09 / 09 / technology / ai - video - deepfake - runway - kling - quiz . html,

    2024

  41. [49]

    De- ferred neural rendering: Image synthesis using neural tex- tures

    Justus Thies, Michael Zollh ¨ofer, and Matthias Nießner. De- ferred neural rendering: Image synthesis using neural tex- tures. Acm Transactions on Graphics (TOG) , 38(4):1–12,

  42. [50]

    Interpretable and trustworthy deepfake detection via dy- namic prototypes

    Loc Trinh, Michael Tsang, Sirisha Rambhatla, and Yan Liu. Interpretable and trustworthy deepfake detection via dy- namic prototypes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1973– 1983, 2021. 2

    1973

  43. [51]

    Attention is all you need

    A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, Ł Kaiser, and I Polosukhin. Attention is all you need. NeurIPS, 2017. 4

    2017

  44. [52]

    Modelscope text-to-video technical report

    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. 1, 2, 7

    2023 arXiv

  45. [53]

    Hififace: 3d shape and semantic prior guided high fidelity face swapping

    Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and 10 Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. Proceedings of the Thirtieth In- ternational Joint Conference on Artificial...

    2021

  46. [54]

    Lavie: High-quality video gener- ation with cascaded latent diffusion models

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

    2023 arXiv

  47. [55]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 22445–22455, 2023. 2

    2023

  48. [56]

    Deepfake video detection using convolutional vision transformer

    Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv preprint arXiv:2102.11126, 2021. 6

    2021 arXiv

  49. [57]

    Tall: Thumbnail layout for deepfake video detection

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

  50. [58]

    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. 5, 6, 8

    2019

  51. [59]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975–11986, 2023. 3

    2023

  52. [60]

    Show-1: Marrying pixel and latent diffusion models for text-to-video generation

    David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Com- puter Vision, pages 1–15, 2024. 7

    2024

  53. [61]

    Avid: Any-length video inpainting with dif- fusion model

    Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with dif- fusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 71...

    2024

  54. [62]

    Istvt: interpretable spatial- temporal video transformer for deepfake detection

    Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: interpretable spatial- temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security , 18: 1335–1348, 2023. 6

    2023

  55. [63]

    Open-sora: Democratizing efficient video production for all, 2024

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2

    2024

  56. [64]

    Fine- grained open-set deepfake detection via unsupervised do- main adaptation

    Xinye Zhou, Hu Han, Shiguang Shan, and Xilin Chen. Fine- grained open-set deepfake detection via unsupervised do- main adaptation. IEEE Transactions on Information Foren- sics and Security, 2024. 2 11

    2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.