pith. sign in

arxiv: 2604.04634 · v1 · submitted 2026-04-06 · 💻 cs.CV · cs.AI

Preserving Forgery Artifacts: AI-Generated Video Detection at Native Scale

Zhengcen Li , Chenyang Jiang , Hang Zhao , Shiyang Zhou , Yunyang Mo , Feng Gao , Fan Yang , Qiben Shan
show 2 more authors
Shaocong Wu Jingyong Su
This is my paper

Pith reviewed 2026-05-10 20:16 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords AI-generated video detectionforgery artifactsnative scale processingVision Transformersynthetic mediahigh-frequency artifactsspatiotemporal inconsistencies
0
0 comments X

The pith

Processing videos at native variable resolutions and durations preserves high-frequency forgery artifacts lost in resizing, enabling better AI-generated video detection.

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

Standard detection methods resize and crop videos to fixed sizes, which discards the subtle high-frequency details and timing inconsistencies that mark synthetic content. This paper curates a dataset of over 140,000 videos from 15 current generators plus a new Magic Videos benchmark, then applies a vision transformer that accepts each video at its original resolution and length without alteration. If the central claim holds, detectors could analyze real-world uploads directly and retain the very clues that resizing erases, raising the bar for reliable identification of sophisticated fakes. The work positions native-scale handling as a core requirement rather than an optional refinement.

Core claim

The authors claim that a detection framework built on the Qwen2.5-VL Vision Transformer, which operates natively at variable spatial resolutions and temporal durations, preserves the high-frequency artifacts and spatiotemporal inconsistencies typically lost during conventional preprocessing and thereby achieves superior performance across multiple benchmarks on a new large-scale dataset.

What carries the argument

Native-scale operation of the vision transformer, which accepts input videos at their original resolutions and durations without fixed resizing or cropping.

Load-bearing premise

That the performance advantage comes specifically from retaining artifacts at native scale rather than from other differences in model choice or dataset composition.

What would settle it

Apply the identical model to the Magic Videos benchmark after standard fixed-resolution resizing and obtain equal or higher accuracy than the native-scale version.

Figures

Figures reproduced from arXiv: 2604.04634 by Chenyang Jiang, Fan Yang, Feng Gao, Hang Zhao, Jingyong Su, Qiben Shan, Shaocong Wu, Shiyang Zhou, Yunyang Mo, Zhengcen Li.

Figure 1
Figure 1. Figure 1: Resolution mismatch and generator quality strongly affect cross-generator video detection. Left: Detectors trained on 720p videos (top) and on lower-resolution videos (<720p; bottom) both exhibit a pronounced performance drop when evaluated at a spatial resolution different from that used during training. Right: We observe a strong positive correlation between generator quality (VBench score) and cross-val… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the data generation pipeline and the proposed detection framework. Left: We curate high-quality captions from real videos and refine them into prompts for state￾of-the-art text-to-video generators, producing realistic synthetic videos for training and evaluation. Right: Our detector supports variable spatial resolutions and temporal lengths. It avoids fixed-size resizing/cropping and applies 3D… view at source ↗
Figure 3
Figure 3. Figure 3: Robustness on MovieGen under compression and spatial perturbations (relative ACC). Perturbation methods include JPEG compression, H264 encoding, spatial resizing and crop￾ping. curacy peaking at 89.92 when using resolutions up to 720p. This confirms our hypothesis that main￾taining aspect ratio and processing at higher resolutions are critical for capturing subtle, pixel-level forgery artifacts. Ablation S… view at source ↗
Figure 4
Figure 4. Figure 4: MDS visualization of generator similarity induced by cross-model detection perfor￾mance.. Model similarity is based on pairwise detection accuracy. Model Wan21 Hun￾yuan Kling Sora Gen-3 Rep￾Video JimengLuma Mira Pika Open Sora STIV Caus Vid VCraf ter-2 ADiff -V2 AVG Wan21 99.7 97.4 93.1 97.1 98.4 88.7 96.8 97.5 33.8 91.3 44.7 88.3 97.7 92.1 94.2 87.38 Hunyuan 93.9 99.4 81.3 92.9 92.0 69.7 89.7 91.4 38.5 84… view at source ↗
Figure 5
Figure 5. Figure 5: Video Visualization from Magic Video Benchmark. From left to right, each column denotes videos from real sources, seaweed, seedance, and wan2.1 [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Saliency Analysis. Saliency maps of our model on AI-generated video samples. Figures 5 and 7 to 9 present a selection of video samples from our dataset, with Figures 3–5 offering detailed visualizations along with their corresponding generative prompts. As illustrated in these figures, the videos in Magic Videos benchmark exhibit high visual quality, characterized by aesthetic appeal, rich motion, and dive… view at source ↗
Figure 7
Figure 7. Figure 7: Video Visualization from Magic Video Benchmark. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Video Visualization from Magic Video Benchmark. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Video Visualization from Magic Video Benchmark. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
read the original abstract

The rapid advancement of video generation models has enabled the creation of highly realistic synthetic media, raising significant societal concerns regarding the spread of misinformation. However, current detection methods suffer from critical limitations. They rely on preprocessing operations like fixed-resolution resizing and cropping. These operations not only discard subtle, high-frequency forgery traces but also cause spatial distortion and significant information loss. Furthermore, existing methods are often trained and evaluated on outdated datasets that fail to capture the sophistication of modern generative models. To address these challenges, we introduce a comprehensive dataset and a novel detection framework. First, we curate a large-scale dataset of over 140K videos from 15 state-of-the-art open-source and commercial generators, along with Magic Videos benchmark designed specifically for evaluating ultra-realistic synthetic content. In addition, we propose a novel detection framework built on the Qwen2.5-VL Vision Transformer, which operates natively at variable spatial resolutions and temporal durations. This native-scale approach effectively preserves the high-frequency artifacts and spatiotemporal inconsistencies typically lost during conventional preprocessing. Extensive experiments demonstrate that our method achieves superior performance across multiple benchmarks, underscoring the critical importance of native-scale processing and establishing a robust new baseline for AI-generated video detection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript claims that current AI-generated video detectors are limited by fixed-resolution resizing and cropping, which discard high-frequency forgery artifacts and spatiotemporal inconsistencies. It introduces a new dataset of over 140K videos from 15 modern generators plus the Magic Videos benchmark, and proposes a detection framework based on the Qwen2.5-VL Vision Transformer that processes videos natively at variable resolutions and durations without preprocessing. The authors assert that this native-scale approach preserves forgery traces and delivers superior performance across benchmarks, establishing a new baseline.

Significance. If the empirical claims are substantiated with controlled experiments, the work would be significant for demonstrating the value of native-scale processing in preserving detectable artifacts from state-of-the-art generators and for supplying a large modern benchmark that could serve as a reference for future detectors.

major comments (3)
  1. [Abstract] Abstract: the claim that the method 'achieves superior performance across multiple benchmarks' is unsupported by any quantitative metrics, baseline comparisons, error bars, or statistical tests, which is load-bearing for the central assertion that native-scale processing drives the gains.
  2. [Method] Method section: the description of the Qwen2.5-VL framework does not specify the detection procedure (zero-shot prompting, added classification head, fine-tuning details, or loss function), preventing attribution of results to native-scale input rather than training choices.
  3. [Experiments] Experiments section: no ablation is reported that trains the identical Qwen2.5-VL backbone under both native variable-resolution and fixed-resolution regimes on the new 140K dataset, so performance deltas cannot be isolated from dataset shift or model selection.
minor comments (2)
  1. [Abstract] The abstract should include at least one concrete performance number (e.g., AUC or accuracy) to allow readers to gauge the magnitude of improvement.
  2. [Dataset] Ensure the dataset curation details (train/test splits, generator versions, and video durations) are tabulated for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough and insightful review. The comments have helped us improve the clarity and rigor of our work. Below, we provide point-by-point responses to the major comments and indicate the revisions made to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the method 'achieves superior performance across multiple benchmarks' is unsupported by any quantitative metrics, baseline comparisons, error bars, or statistical tests, which is load-bearing for the central assertion that native-scale processing drives the gains.

    Authors: We agree with the referee that the abstract should be supported by quantitative evidence. The original abstract was kept concise, but this omitted key details. In the revision, we have expanded the abstract to report specific performance metrics from our experiments, including accuracy, precision, and AUC values with baseline comparisons. Additionally, we have included error bars and statistical significance tests in the experiments section to support the claims of superior performance. This revision directly addresses the load-bearing nature of the assertion. revision: yes

  2. Referee: [Method] Method section: the description of the Qwen2.5-VL framework does not specify the detection procedure (zero-shot prompting, added classification head, fine-tuning details, or loss function), preventing attribution of results to native-scale input rather than training choices.

    Authors: We appreciate this observation, as the method details are crucial for reproducibility and attribution. The original description focused on the native-scale aspect but lacked specifics on the training setup. We have revised the method section to explicitly describe the detection procedure: a classification head is added to the Qwen2.5-VL model, which is fine-tuned using binary cross-entropy loss. Full details on hyperparameters, optimization, and the absence of zero-shot methods are now provided. This ensures that the performance gains can be attributed to the native-scale input processing. revision: yes

  3. Referee: [Experiments] Experiments section: no ablation is reported that trains the identical Qwen2.5-VL backbone under both native variable-resolution and fixed-resolution regimes on the new 140K dataset, so performance deltas cannot be isolated from dataset shift or model selection.

    Authors: We concur that a controlled ablation is essential to isolate the benefits of native-scale processing from other factors. We have conducted this ablation by training the identical Qwen2.5-VL backbone on the 140K dataset under native variable-resolution conditions versus a fixed-resolution regime. The revised experiments section now includes these results, demonstrating that the native-scale variant achieves superior detection performance. This helps confirm that the improvements stem from preserving forgery artifacts rather than dataset or model variations. revision: yes

Circularity Check

0 steps flagged

No circularity; purely empirical claims with no derivations or self-referential reductions

full rationale

The manuscript introduces a new 140K-video dataset from 15 generators plus the Magic Videos benchmark, then applies the unmodified Qwen2.5-VL model at native variable resolutions and durations. No equations, ansatzes, fitted parameters, or predictions appear in the provided text. The central claim—that native-scale processing preserves high-frequency artifacts—is advanced solely via experimental comparisons, not by any derivation that reduces to its own inputs by construction. Self-citations (if present) are not load-bearing; the argument rests on new data and direct model application rather than any uniqueness theorem or prior result from the same authors. This is a standard empirical contribution whose validity can be checked against external benchmarks without circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities beyond reliance on the pre-trained Qwen2.5-VL model and choices in dataset curation; insufficient detail to enumerate further.

pith-pipeline@v0.9.0 · 5541 in / 1177 out tokens · 75372 ms · 2026-05-10T20:16:39.678345+00:00 · methodology

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

19 extracted references · 19 canonical work pages · 1 internal anchor

  1. [1]

    URLhttps://arxiv.org/abs/2412. 19437. 11 Published as a conference paper at ICLR 2026 Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim Alabdulmohsin, Avital Oliver, Piotr Padlewski, Alexey Gritsenko, Mario Luˇci´c, and Neil Houlsby. Patch n’ pack: N...

    work page 2026

  2. [2]

    Vbench++: Comprehensive and versatile benchmark suite for video generative models,

    12 Published as a conference paper at ICLR 2026 Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chan- paisit, Chenyang Si, Yuming Jiang, Yaohui Wang, Xinyuan Chen, Ying-Cong Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Vbench++: Comprehensive and versatile benchmark suite for video generative models,

    work page 2026

  3. [3]

    Stiv: Scalable text and image conditioned video generation,

    13 Published as a conference paper at ICLR 2026 Zongyu Lin, Wei Liu, Chen Chen, Jiasen Lu, Wenze Hu, Tsu-Jui Fu, Jesse Allardice, Zhengfeng Lai, Liangchen Song, Bowen Zhang, Cha Chen, Yiran Fei, Yifan Jiang, Lezhi Li, Yizhou Sun, Kai-Wei Chang, and Yinfei Yang. Stiv: Scalable text and image conditioned video generation,

    work page 2026

  4. [4]

    Towards universal fake image detectors that generalize across generative models

    14 Published as a conference paper at ICLR 2026 Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. InCVPR, pp. 24480–24489,

    work page 2026

  5. [5]

    Roformer: En- hanced transformer with rotary position embedding,

    15 Published as a conference paper at ICLR 2026 Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: En- hanced transformer with rotary position embedding,

    work page 2026

  6. [6]

    H.; Yan, H.; Liu, J.-W.; Zhang, C.; Feng, J.; and Shou, M

    doi: 10.48550/arXiv.2311.16498. URLhttp: //arxiv.org/abs/2311.16498. arXiv:2311.16498 [cs]. Zhiyuan Yan, Yong Zhang, Xinhang Yuan, Siwei Lyu, and Baoyuan Wu. Deepfakebench: A com- prehensive benchmark of deepfake detection. InNeurIPS,

    work page doi:10.48550/arxiv.2311.16498

  7. [7]

    Qwen2.5 Technical Report

    16 Published as a conference paper at ICLR 2026 An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Q...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  8. [8]

    • Section B: Cross-validation experiment

    17 Published as a conference paper at ICLR 2026 APPENDIX This appendix provides a detailed analysis of our dataset, implementation details, additional experi- mental results, and visualizations: • Section A: Data distribution and analysis of our dataset. • Section B: Cross-validation experiment. • Section C: Additional implementation details for both our ...

    work page 2026

  9. [9]

    24.12 private 4720 854x480 30 150 5.0s HunyuanVideo (Kong et al., 2024)24.12 open-source 4725 1280x720 24 129 5.4s Gen-3 (Germanidis,

    work page 2024

  10. [10]

    24.05 private 6214 1280x720 8 96 12.0s OpenSora V1.1 (Zheng et al., 2024)24.04 open-source 4720 424x240 8 64 8.0s Mira (Ju et al.,

    work page 2024

  11. [11]

    23.11 private 4715 1280x720 24 72 3.0s AnimateDiff-V2 (Guo et al., 2024)23.09 open-source 4715 512x512 8 16 2.0s Overall Fake - - 70,692 240-720p 6-60 16-256 1-12s Table 8:Statistics of real and synthetic videos in the proposed training set. Model / Video Split Videos Resolution FPS Frame Duration Movie Gen (Polyak et al., 2024)validation (fake) 1003 1920...

    work page 2024

  12. [12]

    A.1 TRAININGSET

    215 960×540 25 204 8.2s Table 9:Statistics of real and synthetic videos in the proposed validation and Magic Videos Benchmark. A.1 TRAININGSET. Table 8 provides a comprehensive summary of the training dataset used in our work. Previous research has emphasized the critical importance of dataset quality and diversity in training robust detectors (Rajan et a...

    work page 2025

  13. [13]

    The models differ significantly in training methodology, data scale, output resolution, and video duration, contributing to a richly diverse training set

    It includes a wide range of model types in terms of availabil- ity (i.e., open-source, open-report, and private) and architecture (e.g., Diffusion U-Net, DiT-based, auto-regressive models, and others with undisclosed architectures). The models differ significantly in training methodology, data scale, output resolution, and video duration, contributing to ...

    work page 2017

  14. [14]

    These videos are pro- vided at resolutions up to 1080p to ensure both high fidelity and content diversity

    and Pexels(pexels, 2024)—covering a diverse range of common scenes such as landscapes, architecture, human subjects, and news footage. These videos are pro- vided at resolutions up to 1080p to ensure both high fidelity and content diversity. For evaluation, real and generated videos are matched into balanced subsets, allowing for the computation of accu- ...

    work page 2024

  15. [15]

    to produce a 2D spatial representation of the generative models. This visualization aids in understanding the architectural relationships and clustering patterns among the models, offering insights into how architectural similarity correlates with cross-detection performance. Impact of Generation Quality.In addition to architecture,M[i, j]is also influenc...

    work page 2023

  16. [16]

    as a proxy for generation quality. To assess the relationship between generation quality and detection effective- ness, we compute Pearson correlation coefficients (ρ) between the benchmark quality scores and corresponding detection accuracies. B.2 CROSS-VALIDATIONRESULTS Cross Validation.As discussed above, we use the cross-validation matrixMto evaluate ...

    work page 2020

  17. [17]

    and VideoCrafterV2(Chen et al., 2024a), while autoregressive-based models, such as (Yin et al., 2025), appear more distant from the rest. This map- ping also informs a diverse training set selection of generative models, we could combine the cross validation accuracy and similarity to construct a high-quality and diverse dataset for data-efficient trainin...

    work page 2025

  18. [18]

    Since the cross-validation data is directly sampled from VBench’s evaluation set, the VBench scores pro- vide an accurate proxy for the generation quality of each subset

    The results are visualized in Fig.1 of our main paper. Since the cross-validation data is directly sampled from VBench’s evaluation set, the VBench scores pro- vide an accurate proxy for the generation quality of each subset. Across 14 models (excluding CausVid, which features a fundamentally different model structure and training paradigm), we com- pute ...

    work page 2026

  19. [19]

    The results confirm that our native-resolution framework effectively captures two key types of features crucial for AIGC detection. (1) Low-level Artifacts: In billboard scenes, the model focuses on fine details such as distorted text rendering and unnatural edge transitions that are often lost during resolution downsampling. These high-frequency artifact...

    work page 2026