Pith. sign in

REVIEW 3 major objections 5 minor 56 references

AI-generated video attribution works better as retrieval than classification: matching generator-specific fingerprints in a reference database outperforms trained classifiers, reaching 84.6% Rank-1 accuracy and 91.0% detection accuracy on t

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 16:30 UTC pith:BXMUB2ZL

load-bearing objection A training-free retrieval pipeline for AI-video attribution with a genuinely new adaptive color transform and strong-looking numbers, but the evaluation may leak gallery videos into the query set and it skips the closest prior work — worth refereeing, not yet trustable as SOTA. the 3 major comments →

arxiv 2607.28955 v1 pith:BXMUB2ZL submitted 2026-07-31 cs.CV cs.AI

Retrieval-Driven Training-Free AI-Generated Video Attribution

Renxi Cheng , Chaolei Han , Jie Gui , Hongsong Wang This is my paper
classification cs.CV cs.AI
keywords AI-generated video attributiontraining-freeinstance retrievalgenerative fingerprintcolor transformationquantized residualvideo forensicsopen-set attribution
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that attributing an AI-generated video to the specific model that created it should be treated as a retrieval problem, not a classification problem. A query video is matched against a reference database of videos from known generators by comparing generator-specific fingerprints. To expose these fingerprints, the authors design a training-free pipeline that transforms frames into an adaptive color space, computes multi-scale quantized residuals that amplify subtle artifacts, and fuses spatial, temporal, and original RGB streams into a video-level representation. A standard pre-trained video encoder embeds these representations, and cosine similarity over the database yields both source attribution and real-vs-fake detection. On GenVidBench, the method reports 84.6% Rank-1 attribution accuracy and 78.3% mAP in the 100-shot setting, surpassing trained baselines by wide margins while requiring no task-specific training.

Core claim

The central claim is that modern video generators leave stable, model-specific low-level artifacts that survive across videos and can serve as fingerprints for source attribution. The paper shows that these fingerprints are best exposed not in raw RGB frames but after an adaptive orthogonal color transformation that decorrelates channels along the direction of maximal high-frequency energy, followed by multi-scale channel-asymmetric quantization whose residuals amplify fine-grained generation traces. Weighted fusion of spatial residual, temporal residual, and original RGB streams produces a representation that separates eight generators and real videos well enough that simple nearest-neighbo

What carries the argument

The adapted orthogonal color transformation (AOCT): a per-frame 3x3 orthogonal matrix, with one axis constrained to be a non-negative simplex vector dominated by the green channel, chosen to maximize the energy of the high-frequency residual; it has a closed-form solution via the top eigenvector of the residual covariance matrix followed by Gram-Schmidt orthogonalization. Its role is to rotate RGB channels into a coordinate system where generator artifacts are more visible before quantization. The multi-scale quantized residual (MSQR) then quantizes each channel at four scales with asymmetric factors and subtracts the reconstructed frame from the original, yielding residuals that amplify sub

Load-bearing premise

The method assumes that a few reference videos per generator capture that generator's fingerprint well enough that any new video from the same generator will be closer to those references than to any other generator's; this breaks if reference videos are not representative of the generator's output distribution or if query videos overlap with the reference database.

What would settle it

Build a 100-shot database from videos sampled from a single prompt per generator on GenVidBench, then query with videos generated from very different prompts; if Rank-1 accuracy drops to near chance, the fingerprint is content-dependent rather than generator-specific. More directly, register references from one random split of the benchmark and query on a disjoint split, reporting per-split variance, to determine whether the reported 84.6% Rank-1 reflects true separation or database/query overlap.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • New generators can be added to attribution simply by registering reference videos, with no retraining.
  • The same representation doubles as a real-vs-fake detector, since real videos remain dissimilar to all registered generator fingerprints.
  • The method stays robust under H.264 compression and central cropping, suggesting the fingerprints are stable artifacts rather than superficial cues.
  • It is computationally cheaper than training-based attribution methods, requiring only database construction and similarity search.
  • The retrieval formulation naturally supports few-shot attribution, maintaining 52.0% mAP even with a single reference video per source.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the fingerprints are truly model-specific and content-independent, the same retrieval-plus-residual pipeline could transfer to image or audio attribution by swapping the encoder and residual operators, though the temporal stream would need adaptation.
  • The claim 'training-free' applies to attribution training, not to the pre-trained video encoder; the reported rankings could shift if a different encoder were used, since the fingerprint representation is encoder-dependent.
  • Because reference sampling is not specified, a fair evaluation requires disjoint reference/query splits; the reported margins could shrink if references are drawn from a narrow set of prompts while queries cover diverse content.
  • The robustness to compression suggests fingerprints live in structural residuals; a stress test that adversarially removes these residuals (e.g., targeted denoising) would reveal whether the attributions are truly tied to generation artifacts.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a training-free, retrieval-based framework for attributing AI-generated videos to their source generators. The method transforms video frames with an adaptive orthogonal color transformation (AOCT), computes multi-scale quantized residuals (MSQR), and fuses spatial, temporal, and RGB streams before extracting features with a pre-trained R3D-18 backbone. Attribution is performed by cosine-similarity retrieval over a reference database of registered videos. The authors report state-of-the-art results on GenVidBench: 84.6% Rank-1 and 78.3% mAP for 100-shot attribution and 91.0% average detection accuracy, together with ablations and robustness/efficiency analyses.

Significance. If the empirical claims hold, the paper makes a useful contribution: it reformulates video attribution as an open-set retrieval problem, avoids training a source classifier, and supports incremental addition of new generators by registering reference samples. The proposed residual pipeline is well motivated, the ablations are extensive, and the authors provide code and an efficiency comparison. However, the headline results rest on two protocol issues that must be resolved before the claims can be accepted: the query/gallery split is not specified, and key hyperparameters appear to be selected on the same benchmark used for reporting.

major comments (3)
  1. [Sec. 4.1 / Implementation Details] The retrieval protocol is underspecified. The database is defined as 'registering 1, 10, and 100 video features per source,' and queries 'undergo the same process, and cosine similarity is computed against all stored features,' but the paper never states that registered videos are excluded from the query set. If a query video is also in the gallery, its self-match has cosine similarity 1, which inflates Rank-1, mAP, and the detection accuracy in Tables 1-2. This is the central evidence for the 'outperforms all competing methods' claim. Please specify the exact disjoint split between gallery and query videos and rerun the evaluation under that protocol.
  2. [Sec. 4.3 / Implementation Details] Several free hyperparameters appear to be chosen on the test benchmark. Tables 5, 6, and 9 sweep the fusion weights and quantization scales and report the best mAP; the final numbers (e.g., 78.3% mAP, 84.6% Rank-1) use these selected values. Without a validation split or a nested selection protocol, the reported performance is an optimistic selection result rather than a fixed-pipeline result. Please define the validation procedure used to choose the scales and weights, and report test performance only after the choices are fixed.
  3. [Sec. 3.1, Eqs. (1)-(9)] The AOCT derivation is claimed to have a closed-form solution, but the proof is deferred to the Supplementary Material, which is not included with the manuscript. The optimization in Eq. (1) is constrained by an orthogonality condition and a simplex constraint with a threshold on one row; the proposed solution computes the top eigenvector of Sigma, applies a projection Pi_C, and then Gram-Schmidt orthogonalization. The projection Pi_C and threshold tau are not defined, and it is not obvious that this procedure solves the stated constrained problem. Since AOCT is the first stage and contributes about 10.9% mAP in the ablation, please provide the full proof and precise definitions in the main text or an included supplement.
minor comments (5)
  1. [Abstract] The phrase 'with a Rank-1 accuracy of 20.5% and a mean Average Precision of 16.6%' should read 'with improvements of 20.5 percentage points in Rank-1 and 16.6 percentage points in mAP.'
  2. [Fig. 1] The label 'RetrivalModel' is a typo for 'Retrieval Model.'
  3. [Sec. 3.2 / 3.3] The module is called MSQR in the text but 'MSQG' appears in Sec. 3.3; please standardize the abbreviation.
  4. [Sec. 4.2] The description of training DeMamba and UNITE says they are trained on 'randomly selected videos from subsets of GenVidBench' and evaluated on 'all videos'; this makes the comparison to retrieval-based methods unclear. Please report the exact training/evaluation split for these baselines as well.
  5. [Sec. 4.1] For the Real category, the paper combines HD-VG and Vript into one class. Since these are two distinct real-video sources with potentially different characteristics, it would be helpful to report per-source detection accuracy separately or justify the pooling.

Circularity Check

1 steps flagged

No definitional derivation-circle in the fingerprint pipeline, but the headline numbers are partly in-sample: fusion weights and quantization scales are selected by maximizing mAP on GenVidBench, and the query/gallery split is not documented.

specific steps
  1. fitted input called prediction [Sec. 4.3 'Analysis of Fusion Weights' / 'Analysis of Quantization Scales'; final results in Sec. 4.2 'Attribution Performance']
    "We subsequently set w_temporal : w_spatial to 1:2, and vary the RGB weight wRGB from 0.05 to 0.30, maintaining the sum of w_spatial, w_temporal and wRGB as 1. It is evident in Tab. 6 that our approach is also robust to different RGB weights wRGB, and the RGB weight wRGB of 0.10 facilitates the approach in achieving the optimal mAP of 78.3%."

    The reported headline 78.3% mAP (and the Rank-1 numbers in Table 1) are produced with constants explicitly selected by maximizing mAP on GenVidBench: the fusion weights in Tabs. 5-6 and quantization scales in Tab. 9 are chosen using the same benchmark on which the final numbers are reported. Thus the evaluation is not an independent prediction of a fixed training-free pipeline; the reported metric is the in-sample optimum of a hyperparameter search. This is a fitted parameter being presented as a predictive result, even though the feature-extraction equations themselves are not label-fitted.

full rationale

The core method is a self-contained feature construction: AOCT is a closed-form per-frame transform, MSQR is a fixed quantization-residual operation, and temporal aggregation is a weighted sum of spatial, temporal, and RGB streams; none of these equations is defined in terms of GenVidBench labels or the final ranks. The R3D-18 backbone is external (pretrained on Kinetics-400), and the benchmark is external, so the central representation is not simply a renamed version of the target. The only self-citation, LOTA [44], appears as a baseline and is not load-bearing for the method; no uniqueness theorem or ansatz is imported from it. The main circularity concern is empirical rather than derivational: Sec. 4.3 selects fusion weights and quantization factors by maximizing mAP on GenVidBench, then Sec. 4.2 reports those same numbers as the achieved result. Additionally, Sec. 4.1 never states that registered gallery videos are excluded from the query set; if overlap exists, top-1 self-matches would inflate Rank-1/mAP/detection by construction. That overlap condition is not demonstrated in the text, so I treat it as a protocol risk rather than a proven derivation-circle. Overall, the pipeline derivation is not circular, but the headline quantitative claim is partly in-sample, giving a moderate partial-circularity score.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on four assumptions: the pre-trained feature space is discriminative for generator identity, generator artifacts are stable and separable, the benchmark protocol is valid and disjoint, and the AOCT closed-form solution is correct as stated. The free parameters are the quantization scales, fusion weights, AOCT threshold, and scale/frame counts; all are tuned or chosen on the same benchmark. The paper introduces no new physical or conceptual entities beyond a descriptive 'fingerprint' label for computed residual features.

free parameters (4)
  • Quantization scale vectors S^(i) = channel 1: [0.50, 1.20, 1.90, 2.60]; channels 2/3: [0.40, 1.00, 1.60, 2.20]
    Hand-set in Sec. 4.1 and justified via ablation on GenVidBench in Table 9; selecting the best of many configurations on the evaluation benchmark makes the headline mAP a fitted quantity.
  • Fusion weights (w_s, w_t, w_rgb) = 0.6, 0.3, 0.1
    Chosen by sweeping temporal and RGB weights (Tables 5-6) on the same benchmark; no held-out validation is described.
  • AOCT threshold tau and projection Pi_C = unspecified
    Eq. 4 imposes m1,G >= tau but tau is never reported; the projection in Eq. 8 and the claimed closed-form solution depend on it.
  • Number of quantization scales n and frame count T = n=4, T=64
    Ablation shows scale cardinality peaked at four (Table 9); T=64 is an implementation choice. These are engineered parameters that shape the pipeline.
axioms (4)
  • domain assumption Pre-trained R3D-18 on Kinetics-400 provides a feature space in which cosine similarity of the fused residual representation reflects generator identity.
    Section 4.1 uses the Kinetics-400 backbone as feature extractor with no task-specific training or verification of the metric property.
  • domain assumption Generator-specific artifacts are stable within a generator and separable across the eight GenVidBench generators.
    Section 3 states artifacts are 'closely related to the underlying generation mechanisms'; this motivating premise is tested only on this benchmark.
  • domain assumption GenVidBench labels and the few-shot gallery/query protocol are correct and disjoint.
    Section 4.1 does not specify how reference videos are sampled or whether they overlap the query set; the reported numbers depend on this protocol assumption.
  • ad hoc to paper The optimization in Eq. 1 under constraints has the claimed closed-form solution.
    Text says proof is in Supplementary Material, which is not included; main text uses Eq. 7 v_max(Sigma) and Gram-Schmidt without presenting the proof.

pith-pipeline@v1.3.0-daily-deepseek · 17367 in / 15348 out tokens · 131882 ms · 2026-08-03T16:30:47.220292+00:00 · methodology

0 comments
read the original abstract

AI-generated videos are becoming increasingly realistic and difficult to distinguish from authentic ones, which facilitates malicious misuse and poses growing threats to cybersecurity and social governance. Attributing AI-generated videos to their specific generative sources is therefore of critical importance for forensic investigation and legal regulation. However, most existing visual attribution methods focus on images and particularly rely on the image generation model, thereby lacking the ability to generalize to large-scale AI-generated video data. To address these limitations, we introduce an training-free AI-generated video attribution paradigm. Specifically, we formulates AI-generated video attribution as an instance retrieval task, and design a generative fingerprint-based pipeline. This pipeline consists of an adapted orthogonal color transformation, multi-scale quantized residual generation, and temporal-semantic aggregation, progressively capturing and integrating artifacts introduced by generative models across video frames. Extensive experiments on the GenVidBench benchmark demonstrate that our method achieves strong performance in both AI-generated video detection and attribution, outperforming existing state-of-the-art methods with a Rank-1 accuracy of 20.5% and a mean Average Precision of 16.6%. The code is at https://github.com/renxi-seu/Video_Attribution.

Figures

Figures reproduced from arXiv: 2607.28955 by Chaolei Han, Hongsong Wang, Jie Gui, Renxi Cheng.

Figure 1
Figure 1. Figure 1: Comparison between existing AI-generated video attri [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our fingerprint-aware representation pipeline. (1) The input video frames are first computed with an error map [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualizations of residual frames from different generative models. We visualize the raw video frames, spatial residual frames, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of video features distribution from real and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of video attribution results on hard cases for AI-generated video detection. We select videos misclassified by other [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 17 linked inside Pith

  1. [1]

    Ai-generated video de- tection via spatio-temporal anomaly learning, 2024

    Jianfa Bai, Man Lin, and Gang Cao. Ai-generated video de- tection via spatio-temporal anomaly learning, 2024. 2

    2024

  2. [2]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023. 6

    Pith/arXiv arXiv 2023

  3. [3]

    Demamba: Ai- generated video detection on million-scale genvideo bench- mark.arXiv preprint arXiv:2405.19707, 2024

    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. 2, 6, 7, 9

    Pith/arXiv arXiv 2024

  4. [4]

    Videocrafter2: Overcoming data limitations for high-quality video diffusion models.arXiv preprint arXiv:2401.09047, 2024

    Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models.arXiv preprint arXiv:2401.09047, 2024. 6

    Pith/arXiv arXiv 2024

  5. [5]

    Adnet: Delving into generalizable deepfake detection via adaptive expert selection and discrepancy learning.Pattern Recognition, 173:112755, 2026

    Haipeng Chen, Yixin Jia, Zenan Shi, and Dong Zhang. Adnet: Delving into generalizable deepfake detection via adaptive expert selection and discrepancy learning.Pattern Recognition, 173:112755, 2026. 1

    2026

  6. [6]

    A single simple patch is all you need for ai-generated image detection.arXiv preprint arXiv:2402.01123, 2024

    Jiaxuan Chen, Jieteng Yao, and Li Niu. A single simple patch is all you need for ai-generated image detection.arXiv preprint arXiv:2402.01123, 2024. 6, 7, 9

    Pith/arXiv arXiv 2024

  7. [7]

    Seine: Short-to-long video diffu- sion model for generative transition and prediction

    Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffu- sion model for generative transition and prediction. InThe Twelfth International Conference on Learning Representa- tions, 2023. 1

    2023

  8. [8]

    Video Seal: Open and Efficient Video Wa- termarking.arXiv preprint arXiv:2412.09492, 2024

    Pierre Fernandez, Hady Elsahar, I Zeki Yalniz, and Alexan- dre Mourachko. Video Seal: Open and Efficient Video Wa- termarking.arXiv preprint arXiv:2412.09492, 2024. 3

    Pith/arXiv arXiv 2024

  9. [9]

    Leveraging fre- quency analysis for deep fake image recognition

    Joel Frank, Thorsten Eisenhofer, Lea Sch ¨onherr, Asja Fis- cher, Dorothea Kolossa, and Thorsten Holz. Leveraging fre- quency analysis for deep fake image recognition. InInterna- tional Conference on Machine Learning, pages 3247–3258. PMLR, 2020. 1, 2

    2020

  10. [10]

    Pid: Generalized ai-generated images detection with pixelwise decomposition residuals

    Xinghe Fu, Zhiyuan Yan, Zheng Yang, Taiping Yao, Yan- dan Zhao, Shouhong Ding, and Xi Li. Pid: Generalized ai-generated images detection with pixelwise decomposition residuals. InInternational Conference on Machine Learning,

  11. [11]

    Towards discovery and attribution of open-world gan generated images

    Sharath Girish, Saksham Suri, Sai Saketh Rambhatla, and Abhinav Shrivastava. Towards discovery and attribution of open-world gan generated images. InProceedings of the IEEE/CVF international conference on computer vision, pages 14094–14103, 2021. 3

    2021

  12. [12]

    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. InInter- national Conference on Learning Representations, 2024. 1

    2024

  13. [13]

    LTX-Video: Realtime Video Latent Diffusion.arXiv preprint arXiv:2501.00103,

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. LTX-Video: Realtime Video Latent Diffusion.arXiv preprint arXiv:2501.00103,

    Pith/arXiv arXiv

  14. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 6, 7

    2016

  15. [15]

    Exposing ai-generated videos: A benchmark dataset and a local-and-global temporal defect based detec- tion method, 2024

    Peisong He, Leyao Zhu, Jiaxing Li, Shiqi Wang, and Hao- liang Li. Exposing ai-generated videos: A benchmark dataset and a local-and-global temporal defect based detec- tion method, 2024. 2

    2024

  16. [16]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022. 6

    Pith/arXiv arXiv 2022

  17. [17]

    VideoShield: Regulating Diffusion-based Video Generation Models via Watermark- ing

    Runyi Hu, Jie Zhang, Yiming Li, Jiwei Li, Qing Guo, Han Qiu, and Tianwei Zhang. VideoShield: Regulating Diffusion-based Video Generation Models via Watermark- ing. InInternational Conference on Learning Representa- tions, 2025. 3

    2025

  18. [18]

    VideoMark: A Distortion-Free Robust Watermarking Framework for Video Diffusion Models.arXiv preprint arXiv:2504.16359, 2025

    Xuming Hu, Hanqian Li, Jungang Li, Yu Huang, Shuliang Liu, Qi Zheng, Junhao Chen, and Aiwei Liu. VideoMark: A Distortion-Free Robust Watermarking Framework for Video Diffusion Models.arXiv preprint arXiv:2504.16359, 2025. 3

    arXiv 2025

  19. [19]

    Distinguish any fake videos: Unleashing the power of large-scale data and motion features, 2024

    Lichuan Ji, Yingqi Lin, Zhenhua Huang, Yan Han, Xiaogang Xu, Jiafei Wu, Chong Wang, and Zhe Liu. Distinguish any fake videos: Unleashing the power of large-scale data and motion features, 2024. 1, 2

    2024

  20. [20]

    Detecting the Abuse of Genera- tive AI in Cybersecurity Contexts: Challenges, Frameworks, and Solutions.Journal of Data Analysis and Critical Man- agement, 1(03):1–12, 2025

    Gopalakrishna Karamchand. Detecting the Abuse of Genera- tive AI in Cybersecurity Contexts: Challenges, Frameworks, and Solutions.Journal of Data Analysis and Critical Man- agement, 1(03):1–12, 2025. 1

    2025

  21. [21]

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

    Levon Khachatryan, Andranik Movsisyan, Vahram Tade- vosyan, Roberto Henschel, Zhangyang Wang, 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 Computer Vision, pages 15954–15964, 2023. 6

    2023

  22. [22]

    HunyuanVideo: A Systematic Framework For Large Video Generative Models.arXiv preprint arXiv:2412.03603,

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. HunyuanVideo: A Systematic Framework For Large Video Generative Models.arXiv preprint arXiv:2412.03603,

    Pith/arXiv arXiv

  23. [23]

    Roy-Chowdhury

    Rohit Kundu, Vishal Mohanty, Hao Xiong, Shan Jia, Athula Balachandran, and Amit K. Roy-Chowdhury. Saga: Source attribution of generative ai videos, 2025. 2, 3

    2025

  24. [24]

    Roy-Chowdhury

    Rohit Kundu, Hao Xiong, Vishal Mohanty, Athula Balachan- dran, and Amit K. Roy-Chowdhury. Towards a universal synthetic video detector: From face or background manip- ulations to fully ai-generated content, 2025. 2, 6, 7, 9

    2025

  25. [25]

    Exposing deepfake videos by detecting face warping artifacts.arXiv preprint arXiv:1811.00656, 2018

    Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts.arXiv preprint arXiv:1811.00656, 2018. 1, 2

    Pith/arXiv arXiv 2018

  26. [26]

    Structure searchable net- work model for deepfake image detection.Pattern Recogni- tion, 179:113813, 2026

    Zinian Liu, Xiaofeng Wang, Qin Wang, Ningning Bai, Ruidong Han, and Shanmin Pang. Structure searchable net- work model for deepfake image detection.Pattern Recogni- tion, 179:113813, 2026. 1

    2026

  27. [27]

    DVMark: A Deep Multiscale Framework for Video Watermarking.IEEE Transactions on Image Processing, 2023

    Xiyang Luo, Yinxiao Li, Huiwen Chang, Ce Liu, Peyman Milanfar, and Feng Yang. DVMark: A Deep Multiscale Framework for Video Watermarking.IEEE Transactions on Image Processing, 2023. 3

    2023

  28. [28]

    Step-Video-T2V Technical Re- port: The Practice, Challenges, and Future of Video Founda- tion Model.arXiv preprint arXiv:2502.10248, 2025

    Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiao- niu Song, Xing Chen, et al. Step-Video-T2V Technical Re- port: The Practice, Challenges, and Future of Video Founda- tion Model.arXiv preprint arXiv:2502.10248, 2025. 1

    Pith/arXiv arXiv 2025

  29. [29]

    Detecting ai-generated video via frame consistency, 2025

    Long Ma, Zhiyuan Yan, Qinglang Guo, Yong Liao, Haiyang Yu, and Pengyuan Zhou. Detecting ai-generated video via frame consistency, 2025. 1, 2

    2025

  30. [30]

    Exploring transferable inconsistencies with re- gional guidance for reference-based deepfake detection.Pat- tern Recognition, 179:113795, 2026

    Liyue Ming, Peisong He, Beibei Li, Haoliang Li, and Xing- hao Jiang. Exploring transferable inconsistencies with re- gional guidance for reference-based deepfake detection.Pat- tern Recognition, 179:113795, 2026. 1

    2026

  31. [31]

    Genvidbench: A challenging benchmark for detecting ai-generated video, 2025

    Zhenliang Ni, Qiangyu Yan, Mouxiao Huang, Tianning Yuan, Yehui Tang, Hailin Hu, Xinghao Chen, and Yunhe Wang. Genvidbench: A challenging benchmark for detecting ai-generated video, 2025. 5

    2025

  32. [32]

    VGMShield: Mitigating Misuse of Video Generative Mod- els.arXiv preprint arXiv:2402.13126, 2024

    Yan Pang, Baicheng Chen, Yang Zhang, and Tianhao Wang. VGMShield: Mitigating Misuse of Video Generative Mod- els.arXiv preprint arXiv:2402.13126, 2024. 2, 3

    Pith/arXiv arXiv 2024

  33. [33]

    Towards Understanding Unsafe Video Generation.arXiv preprint arXiv:2407.12581, 2024

    Yan Pang, Aiping Xiong, Yang Zhang, and Tianhao Wang. Towards Understanding Unsafe Video Generation.arXiv preprint arXiv:2407.12581, 2024. 1

    Pith/arXiv arXiv 2024

  34. [34]

    pika. Pika. Online, 2024. Available:https://pika. art/home. 6

    2024

  35. [35]

    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

    2019

  36. [36]

    Recurrent convo- lutional strategies for face manipulation detection in videos

    Ekraam Sabir, Jiaxin Cheng, Ayush Jaiswal, Wael AbdAl- mageed, Iacopo Masi, and Prem Natarajan. Recurrent convo- lutional strategies for face manipulation detection in videos. Interfaces (GUI), 3(1):80–87, 2019. 2

    2019

  37. [37]

    AI-generated footage falsely used to show ’af- termath’ of Philippines earthquake.https://spotlight.ebu.ch/,

    Spotlight. AI-generated footage falsely used to show ’af- termath’ of Philippines earthquake.https://spotlight.ebu.ch/,

  38. [38]

    Morph studio.https : / / www

    Morph Studio. Morph studio.https : / / www . morphstudio.com/, 2024. 1

    2024

  39. [39]

    A closer look at spatiotemporal convolutions for action recognition

    Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion, 2018. 6, 7

    2018

  40. [40]

    Vript: A video is worth thousands of words

    Shanghai Jiao Tong University. Vript: A video is worth thousands of words. Online, 2024. Available:https: //github.com/mutonix/Vript. 6

    2024

  41. [41]

    Beyond deepfake images: Detecting ai-generated videos

    Danial Samadi Vahdati, Tai D Nguyen, Aref Azizpour, and Matthew C Stamm. Beyond deepfake images: Detecting ai-generated videos. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 4397–4408, 2024. 2, 3

    2024

  42. [42]

    Wan: Open and Advanced Large-Scale Video Generative Models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and Advanced Large-Scale Video Generative Models.arXiv preprint arXiv:2503.20314, 2025. 1

    Pith/arXiv arXiv 2025

  43. [43]

    Swift: Sliding window reconstruction for few-shot training-free generated video attribution, 2026

    Chao Wang, Zijin Yang, Yaofei Wang, Yuang Qi, Weim- ing Zhang, Nenghai Yu, and Kejiang Chen. Swift: Sliding window reconstruction for few-shot training-free generated video attribution, 2026. 3

    2026

  44. [44]

    Lota: Bit-planes guided ai-generated image de- tection

    Hongsong Wang, Renxi Cheng, Yang Zhang, Chaolei Han, and Jie Gui. Lota: Bit-planes guided ai-generated image de- tection. InProceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), pages 17246–17255,

  45. [45]

    Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571, 2023

    Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571, 2023. 6

    Pith/arXiv arXiv 2023

  46. [46]

    Videofactory: Swap at- tention in spatiotemporal diffusions for text-to-video gener- ation.arXiv preprint arXiv:2305.10874, 2023

    Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap at- tention in spatiotemporal diffusions for text-to-video gener- ation.arXiv preprint arXiv:2305.10874, 2023. 6

    Pith/arXiv arXiv 2023

  47. [47]

    Where did i come from? origin attribution of ai-generated images.Advances in neural information pro- cessing systems, 36:74478–74500, 2023

    Zhenting Wang, Chen Chen, Yi Zeng, Lingjuan Lyu, and Shiqing Ma. Where did i come from? origin attribution of ai-generated images.Advances in neural information pro- cessing systems, 36:74478–74500, 2023. 3

    2023

  48. [48]

    Musev: Infinite-length and high fidelity virtual human video genera- tion with visual conditioned parallel denoising.arxiv, 2024

    Zhiqiang Xia, Zhaokang Chen, Bin Wu, Chao Li, Kwok-Wai Hung, Chao Zhan, Yingjie He, and Wenjiang Zhou. Musev: Infinite-length and high fidelity virtual human video genera- tion with visual conditioned parallel denoising.arxiv, 2024. 6

    2024

  49. [49]

    Dynamicrafter: Animating open-domain images with video diffusion priors

    Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamicrafter: Animating open-domain images with video diffusion priors. InEu- ropean Conference on Computer Vision, pages 399–417. Springer, 2024. 1

    2024

  50. [50]

    Video forgery detection with optical flow residuals and spatial-temporal consistency, 2025

    Xi Xue, Kunio Suzuki, Nabarun Goswami, and Takuya Shin- tate. Video forgery detection with optical flow residuals and spatial-temporal consistency, 2025. 1, 2

    2025

  51. [51]

    Progressive open space ex- pansion for open-set model attribution

    Tianyun Yang, Danding Wang, Fan Tang, Xinying Zhao, Juan Cao, and Sheng Tang. Progressive open space ex- pansion for open-set model attribution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15856–15865, 2023. 3

    2023

  52. [52]

    Frequency-aware spatiotemporal transformers for video in- painting detection

    Bingyao Yu, Wanhua Li, Xiu Li, Jiwen Lu, and Jie Zhou. Frequency-aware spatiotemporal transformers for video in- painting detection. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 8188–8197,

  53. [53]

    Mora: Enabling gen- eralist video generation via a multi-agent framework, 2024

    Zhengqing Yuan, Ruoxi Chen, Zhaoxu Li, Haolong Jia, Li- fang He, Chi Wang, and Lichao Sun. Mora: Enabling gen- eralist video generation via a multi-agent framework, 2024. 6

    2024

  54. [54]

    Robust Invisible Video Water- marking with Attention.arXiv preprint arXiv:1909.01285,

    Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust Invisible Video Water- marking with Attention.arXiv preprint arXiv:1909.01285,

    Pith/arXiv arXiv 1909

  55. [55]

    A Novel Deep Video Watermarking Framework with Enhanced Robustness to H

    Yulin Zhang, Jiangqun Ni, Wenkang Su, and Xin Liao. A Novel Deep Video Watermarking Framework with Enhanced Robustness to H. 264/A VC Compression. InProceedings of the 31st ACM International Conference on Multimedia, pages 8095–8104, 2023. 3

    2023

  56. [56]

    Open-Sora: Democratizing Efficient Video Production for All.arXiv preprint arXiv:2412.20404, 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.arXiv preprint arXiv:2412.20404, 2024. 1

    Pith/arXiv arXiv 2024