Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Leveraging Pre-Trained Visual Models for AI-Generated Video Detection

T0 review · 3 major / 6 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 Pre-trained visual features alone can distinguish real from AI-generated video with over 90% average accuracy.

desk verdict The one-to-many protocol leaks real videos into the reference set, so the >90% claim is not supported; the many-to-many results may still be salvageable. read the letter →

arxiv 2507.13224 v1 pith:UHC3E6ZP submitted 2025-07-17 cs.CV

Keerthi Veeramachaneni , Praveen Tirupattur , Amrit Singh Bedi , Mubarak Shah This is my paper
classification cs.CV
keywords AI-generatedvideodetectiontext-to-videogenerationpre-trainedvisualmodelstraining-freelinearprobingSigLIPMAEVID-AIDdataset
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 establishes that generic AI-generated video—not just deepfaked faces—can be detected without training a large detector. Features extracted from frozen pre-trained visual models (SigLIP applied frame-by-frame and averaged, VideoMAE on whole clips) already separate real clips from text-to-video generations well enough that a training-free nearest-distance classifier averages above 90% accuracy on the authors' VID-AID dataset. Adding a single linear layer on the frozen features improves performance further, with only about 1.5K trainable parameters and 4.5K generated videos for training. The same setup also outperforms a 125M-parameter state-space detector on the GenVideo benchmark. A cheap, updateable detector for AI video would matter for flagging misinformation and verifying media provenance.

What carries the argument

The carrying mechanism is the frozen feature space of publicly pre-trained visual encoders. SigLIP, an image-text model trained with a sigmoid loss, is applied frame-by-frame and averaged over time; VideoMAE, a self-supervised masked video model, encodes whole clips. No generator-specific artifact detector is learned. Instead, classification is either Euclidean distance from the test feature to real and generated reference features, or a single trained linear layer on top of the frozen features, which acts as a parameter-efficient probe of the pre-existing separation.

What would settle it

Run the same frozen-feature protocol on a fresh corpus where every generated clip is confirmed by its creator and real clips are diverse random uploads; if per-model F1 scores for labeling closed-source generated clips stay below roughly 30% (as the smallest values in the paper's own many-to-many tables already show for Sora and Veo), the claim of generic above-90% detection fails.

Watch

Extended reading notes

Core claim

The central discovery is that encoders trained only to understand real visual content leave a latent fingerprint of generatedness. Videos from nine text-to-video models—five open-source and four closed-source, including Sora, Veo, Dream Machine, and VideoPoet—are separable from YouTube-VOS real clips in SigLIP and VideoMAE feature space. The paper demonstrates this in two ways: by nearest-distance classification against a reference set, and by training one linear classification layer on frozen features. On VID-AID, a collection of about 10,000 generated and 4,000 real 2-second clips, both variants average above 90% accuracy under one-to-many and many-to-many generalization protocols. The same features also outperform the 125M-parameter DeMamba detector on the GenVideo benchmark while training on a fraction of the data.

Load-bearing premise

The evaluation assumes VID-AID's real videos (all from YouTube-VOS) and scraped social-media clips from closed models are representative of the real and AI-generated videos the detector will meet in practice; if those sources differ in compression or style, the reported accuracies will not carry over.

Editorial extensions

If this is right

  • AI-video detection can be deployed without collecting large training sets or fine-tuning heavy models; a new generator can be handled by adding a few reference samples.
  • Since the detector is a probe on frozen features, it can be updated for new models by retraining only the linear layer, keeping compute low.
  • The method transfers across dataset boundaries: features trained for image understanding (SigLIP) detect video generation, and the same probes beat a video-specific state-space model on GenVideo.
  • Detection difficulty is not uniform: open-source outputs are nearly perfectly separable, while the newest closed-source models like Sora and Veo are the hardest and sometimes near chance in fake-F1, marking the current frontier of the attack-defense race.

Reading between the lines

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

  • The reported separation may be driven partly by resolution, frame rate, and compression artifacts of social-media scraping rather than an intrinsic 'generatedness' signal; a test on verified, uncompressed generated outputs would isolate the true cause.
  • Averaging frame features discards temporal inconsistencies, so fusing per-frame statistics or temporal patch features could push closed-source-model accuracy up—the paper's own future-work suggestion points at spatio-temporal patches.
  • The same frozen-feature recipe should transfer to audio and multimodal generation detection, since any generative model leaves a distributional gap in a pre-trained representation of the corresponding modality.
  • As closed-source generators improve, the reference set must include the newest models; the paper's Veo and Sora results suggest a detector frozen today will age quickly without new reference samples.
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

3 major / 6 minor

Summary. The manuscript proposes a method for detecting AI-generated videos by extracting features from pre-trained visual models (SigLIP and VideoMAE). Two variants are evaluated: a training-free nearest-neighbor distance approach and a training-based linear classifier on the extracted features. The authors introduce VID-AID, a dataset of ~10,000 generated videos from nine text-to-video models and ~4,000 real videos from YouTube-VOS, and report results under one-to-many and many-to-many generalization protocols, claiming detection accuracy above 90% on average. The supplementary material additionally compares the method with DeMamba on the GenVideo benchmark, reporting competitive accuracy with far fewer training data.

Significance. If the central claim holds, the paper offers a practical, parameter-efficient detection method (1.5K trainable parameters) and a useful new dataset for a rapidly evolving problem. The paper's strengths include the public release plans for code, models, and dataset, the cross-benchmark comparison on GenVideo, and the exploration of recent closed-source generation models (Sora, Veo, Dream Machine, VideoPoet), which are largely absent from existing datasets. However, the evaluation protocol contains a data leak that undermines the one-to-many generalization results, and the averaging behind the headline 'above 90%' is not defined.

major comments (3)
  1. [§5.2, Tables 2 and 4] Section 5.2 states that in the one-to-many generalization experiments 'there was no dataset split, and all features from each model were used in different combinations of reference/training and testing.' Because the real-video set (YouTube-VOS) is the same source for reference/training and testing, every real test video is present in the reference/training set: for the training-free approach (Table 2) this yields a distance-zero nearest neighbor, and for the training-based approach (Table 4) the test real videos are directly seen during training. The near-universal F1-Real values above 95% in Tables 2 and 4 are therefore by construction and do not reflect detector quality. These results cannot support the abstract's claim of 'above 90% on average' and must be re-run with disjoint real-video splits (or removed from the paper).
  2. [§5.3, Tables 3 and 5] The 'above 90% on average' claim in the Abstract and Section 5.3 is not backed by a defined aggregation. In the many-to-many results, F1-Fake for the most recent closed-source models is far below 90%: for example, Table 5 (SigLIP, 'Open-source models' training row) reports F1-Fake 25.0 on Veo and 16.2 on Sora, and Table 3 (VideoMAE, 'Open-source models + Sora' row) reports F1-Fake 55.3 on Sora. The paper should specify the exact averaging (macro vs micro, per-class vs per-model) and report a single macro-F1 or balanced accuracy per protocol so the headline claim is verifiable.
  3. [§3 (Dataset)] Section 3 notes that closed-source generated videos were scraped from social media without verification of generation provenance or artifact distribution, while all real videos come from YouTube-VOS. This creates a distribution mismatch with deployment scenarios (e.g., social media platforms). The authors should discuss this selection bias explicitly and report results disaggregated by source or by resolution; the GenVideo comparison in the supplement is a step in this direction but is not part of the main evaluation.
minor comments (6)
  1. [Affiliations and References] There are several typos and formatting issues: 'Univeristy' in affiliations, inconsistent capitalization of 'F1-Score', and the garbled author name in reference [9].
  2. [§5.2] The sentence describing the one-to-many setup ('there was no dataset split...') is confusing and should be rewritten to state the exact composition of reference/training and test sets.
  3. [Figure 1] The caption of Figure 1 includes 'Detection Score' values that are not defined in the text; please specify how these scores were computed.
  4. [§3] Section 3 claims 'a balanced dataset for comparative analysis,' but the ratio of real to generated videos is approximately 1:2.5; adjust the wording or the class balance.
  5. [Supplement, Table 7] In Table 7, the training set size is given as 'OpenSora (1K)' without clarifying whether these are 1,000 videos from VID-AID and whether real videos were included in training; please specify exactly.
  6. [Tables 6 and 7] Tables 6 and 7 use accuracy, while the main tables use F1; the relationship between these metrics should be stated so readers can compare.

Circularity Check

1 steps flagged · score 6.0 of 10

One-to-many evaluation leaks real-video test items into the reference/training set, forcing high F1-real by construction.

  1. fitted input called prediction [Section 5.2 (Implementation Details), one-to-many generalization experiments; results in Tables 2 and 4]
    "In the one-to-many generalization experiments, there was no dataset split, and all features from each model were used in different combinations of reference/training and testing."

    Because no split is made, the YouTube-VOS real videos used to build the reference set for the training-free method, or the training set for the trained linear layer, are the same real videos on which F1-real is later computed. A real test video is therefore either its own nearest neighbor at distance zero in the reference set or a training example already seen by the classifier, so near-perfect F1-real follows from the evaluation protocol itself, not from the proposed features. The tables consistently show F1-real values near 99-100 even when F1-fake for closed-source models is 20-70 (e.g., Veo and VideoPoet rows), which is exactly the signature of this leakage.

full rationale

This paper is an empirical study rather than a mathematical derivation, so most of the chain is not circular: the pre-trained-feature hypothesis is tested on held-out data in the many-to-many protocol, and the GenVideo benchmark in the supplement provides an external, properly split evaluation. The significant circularity is confined to the one-to-many protocol, where the implementation details explicitly state that no dataset split was performed and all features from each model were used in different reference/training and testing combinations. That makes the high F1-real values for the real class a consequence of test items being present in the reference or training set, i.e., a fitted-input-called-prediction / evaluation leak. Because the abstract's headline performance claim is at least partly grounded in these one-to-many VID-AID results, the reported >90% average is partially forced. I therefore assign 6 rather than a lower score; the method retains independent support from the many-to-many and GenVideo experiments.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The central claim rests on two domain assumptions: that pretrained visual features from real content carry transferable signals for AI-video detection, and that the VID-AID dataset is representative. No new entities or fitted constants are introduced; the linear layer weights are standard trainable parameters rather than free parameters that the derivation relies on.

assumptions (2)
  • domain assumption Features from pre-trained visual models (SigLIP, VideoMAE) trained on real content encode sufficient signal to distinguish real vs AI-generated videos.
    This is the central hypothesis, asserted in Section 4.1 and not derived; the entire method depends on it.
  • domain assumption Real videos from YouTube-VOS and scraped closed-source videos are representative of the real and AI-generated video distributions.
    The validity of the reported accuracies depends on this; Section 3 describes the collection but provides no validation of representativeness or label correctness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Pre-Trained Visual Models for AI-Generated Video Detection." pith.science (2026). https://pith.science/paper/UHC3E6ZP

@misc{pith2026250713224,
  author       = {Pith},
  title        = {Pith review of: Leveraging Pre-Trained Visual Models for AI-Generated Video Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UHC3E6ZP}},
  note         = {Machine review of arXiv:2507.13224}
}
read the original abstract

Recent advances in Generative AI (GenAI) have led to significant improvements in the quality of generated visual content. As AI-generated visual content becomes increasingly indistinguishable from real content, the challenge of detecting the generated content becomes critical in combating misinformation, ensuring privacy, and preventing security threats. Although there has been substantial progress in detecting AI-generated images, current methods for video detection are largely focused on deepfakes, which primarily involve human faces. However, the field of video generation has advanced beyond DeepFakes, creating an urgent need for methods capable of detecting AI-generated videos with generic content. To address this gap, we propose a novel approach that leverages pre-trained visual models to distinguish between real and generated videos. The features extracted from these pre-trained models, which have been trained on extensive real visual content, contain inherent signals that can help distinguish real from generated videos. Using these extracted features, we achieve high detection performance without requiring additional model training, and we further improve performance by training a simple linear classification layer on top of the extracted features. We validated our method on a dataset we compiled (VID-AID), which includes around 10,000 AI-generated videos produced by 9 different text-to-video models, along with 4,000 real videos, totaling over 7 hours of video content. Our evaluation shows that our approach achieves high detection accuracy, above 90% on average, underscoring its effectiveness. Upon acceptance, we plan to publicly release the code, the pre-trained models, and our dataset to support ongoing research in this critical area.

Figures

Figures reproduced from arXiv: 2507.13224 by the authors.

Figure 1
Figure 1. Timeline illustrating the advancements in text-to-video [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of SigLIP [38] features of the real videos and the videos from different T2V models in our dataset. 4.2. Training-Free Approach We first address the detection problem using a training-free, distance-based approach. Originally proposed by Ojha et al. [18] to detect generated images, this method relies on fea￾tures extracted from a pre-trained CLIP [22] model. We extend this approach to detect gene… view at source ↗
Figure 4
Figure 4. Overview of the training-based approach, which involves [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Overview of the training-free, distance-based approach [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]
Figure 5
Figure 5. Figure 5: Word Cloud showing the diversity of actors in the videos from our dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Word Cloud showing all the verbs, corresponding to the actions depicted in the videos in our dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Word Cloud showing different scenes captured in the videos from our dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Detecting AI-Generated Video: A Vision-Language Dual-View Survey

    cs.CV 2026-07 conditional novelty 6.0 of 10

    AIGC-V detection should be treated as factual fidelity verification and organized by a four-layer vision-language dual-view taxonomy spanning cues, motion, cross-modal consistency, and world-level reasoning.

Reference graph

Works this paper leans on

41 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    Lumiere: A space- time diffusion model for video generation

    Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Her- rmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A space- time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 3

    arXiv 2024

  2. [2]

    V-jepa: Latent video prediction for visual represen- tation learning

    Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mido Assran, and Nicolas Ballas. V-jepa: Latent video prediction for visual represen- tation learning. 2023. 9

    work page 2023

  3. [3]

    Align your latents: High-resolution video synthesis with la- tent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dock- horn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with la- tent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22563–22575, 2023. 2

    work page 2023

  4. [4]

    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

  5. [5]

    Demamba: Ai- generated video detection on million-scale genvideo bench- mark

    Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, and Huaxiong Li. Demamba: Ai- generated video detection on million-scale genvideo bench- mark. arXiv preprint arXiv:2405.19707 , 2024. 6, 7, 8, 9, 10

    arXiv 2024

  6. [6]

    Audio-visual person-of-interest deep- fake detection

    Davide Cozzolino, Alessandro Pianese, Matthias Nießner, and Luisa Verdoliva. Audio-visual person-of-interest deep- fake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 943–952,

    work page

  7. [7]

    Animatediff: Animate your personalized text- to-image diffusion models without specific tuning

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text- to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representa- tions. 2

    work page

  8. [8]

    Photorealistic video generation with diffusion models

    Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and Jos ´e Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 3 Model Detection Level Testing (real and fake)Sora Gen2 MorphStudio ModelScope Show Lavie Wildscrape Crafter MoonValley HotShot RealAverage CLIP-B-PT* ...

    arXiv 2023

Show all 41 references
  1. [9]

    G ¨uera and E

    D. G ¨uera and E. J. Delp. Deepfake video detection using recurrent neural networks. In 2018 15th IEEE International Conference on Advanced Video and Signal Based Surveil- lance (AVSS), pages 1–6, 2018. 3

    2018

  2. [10]

    Word Cloud showing different scenes captured in the videos from our dataset

    Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Figure 7. Word Cloud showing different scenes captured in the videos from our dataset. Weng, Ying Shan, et al. Animate-a-story: Storytelling with retrieval-augmented video...

    2023 arXiv

  3. [11]

    Animate anyone: Consistent and controllable image- to-video synthesis for character animation

    Li Hu. Animate anyone: Consistent and controllable image- to-video synthesis for character animation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153–8163, 2024. 2

    2024

  4. [12]

    Text2video-zero: Text- to-image diffusion models are zero-shot video generators

    Levon Khachatryan, Andranik Movsisyan, Vahram Tade- vosyan, Roberto Henschel, Wang Zhangyang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text- to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Comput...

    2023

  5. [13]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  6. [14]

    Videopoet: A large language model for zero-shot video gen- eration

    Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vigh- nesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video gen- eration. In Forty-first International Conference on Machine Lear...

    2024

  7. [15]

    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. 3, 5

    2024 arXiv

  8. [16]

    Mao and J

    M. Mao and J. Yang. Exposing deepfake with pixel-wise ar and ppg correlation from faint signals. arXiv preprint, 2021. 3

    2021

  9. [17]

    Expanding language-image pretrained models for gen- eral video recognition

    Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for gen- eral video recognition. In European Conference on Com- puter Vision, pages 1–18. Springer, 2022. 9

    2022

  10. [18]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480– 24489, 2023. 3, 4

    2023

  11. [19]

    Chatgpt (gpt-3.5), 2024

    OpenAI. Chatgpt (gpt-3.5), 2024. Accessed: August 2024. 3

    2024

  12. [20]

    Pytorch: An im- perative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im- perative style, high-performance deep learning library. Ad- vances in neural information processing systems ...

    2019

  13. [21]

    C. Peng, Z. Miao, D. Liu, N. Wang, R. Hu, and X. Gao. Where deepfakes gaze at? spatial–temporal gaze inconsis- tency analysis for video face forgery detection. IEEE Trans- actions on Information Forensics and Security , 19:4507– 4517, 2024. 3

    2024

  14. [22]

    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. In International conference on machine learning, ...

    2021

  15. [23]

    Dream machine

    Dominic Rampas, Thomas Neff, Samarth Sinha, Dan Kon- dratyuk, and Nathan McClean. Dream machine. https: //lumalabs.ai/dream-machine, 2024. 3, 5, 8

    2024

  16. [24]

    Abhishek Sharma, Adams Yu, Ali Razavi, Andeep Toor, Andrew Pierson, and Ankush Gupta. Veo. https:// deepmind.google/technologies/veo/ , 2024. 2, 3, 4, 5, 8

    2024

  17. [25]

    Z. Tong, Y . Song, J. Wang, and L. Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint, 2022. Accessed: Jun. 11,

    2022

  18. [26]

    Visualizing data using t-sne

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579–2605, 2008. 4

    2008

  19. [27]

    Phenaki: Variable length video generation from open domain textual descriptions

    Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kin- dermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Le...

    2022

  20. [28]

    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. 2, 3, 5

    2023 arXiv

  21. [29]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695–8704, 2020. 3

    2020

  22. [30]

    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. 3

    2023

  23. [31]

    Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation

    Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference...

    2023

  24. [32]

    Youtube-vos: A large-scale video object segmentation benchmark

    Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 4, 5

    2018 arXiv

  25. [33]

    Ucf: Uncovering common features for generalizable deep- fake detection

    Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deep- fake detection. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 22412–22423,

  26. [34]

    Cross-modal con- textualized diffusion models for text-guided visual genera- tion and editing

    Ling Yang, Zhilong Zhang, Zhaochen Yu, Jingwei Liu, Minkai Xu, Stefano Ermon, and CUI Bin. Cross-modal con- textualized diffusion models for text-guided visual genera- tion and editing. In The Twelfth International Conference on Learning Representations. 2

  27. [35]

    Y . Yang, Z. Qian, Y . Zhu, and Y . Wu. D3: Scaling up deep- fake detection by learning from discrepancy. arXiv preprint,

  28. [36]

    Nuwa-xl: Diffusion over dif- fusion for extremely long video generation

    Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over dif- fusion for extremely long video generation. In Proceedings of the 61st Annual Meeting of the Association for C...

    2023

  29. [37]

    Zeroscope, 2023

    Zeroscope. Zeroscope, 2023. Accessed: August 2024. 5

    2023

  30. [38]

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer. Sigmoid loss for language image pre-training. arXiv preprint, 2023. 4, 6

    2023

  31. [39]

    Multi-attentional deep- fake detection

    Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep- fake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2185– 2194, 2021. 1

    2021

  32. [40]

    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. 3, 5

    2024

  33. [2024]

    01, 2024

    Accessed: Jun. 01, 2024. 3

    2024

Pith tools

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