pith. machine review for the scientific record. sign in

arxiv: 2604.12307 · v1 · submitted 2026-04-14 · 💻 cs.CV

Recognition: unknown

Boosting Robust AIGI Detection with LoRA-based Pairwise Training

Hao Sun, Qi Zhang, Ruiyang Xia, Xuelong Li, Yaowen Xu, Zhaofan Zou, Zhongjiang He

Pith reviewed 2026-05-10 14:51 UTC · model grok-4.3

classification 💻 cs.CV
keywords AI-generated image detectionrobust detectionLoRA fine-tuningpairwise trainingimage distortionsvisual foundation modelsNTIRE challenge
0
0 comments X

The pith

LoRA-based pairwise training improves AI-generated image detection under severe real-world distortions.

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

The paper targets the gap between clean-dataset performance and real-world failure in AI-generated image detectors, where unpredictable distortions like compression and resizing break standard models. It establishes that finetuning a visual foundation model with deliberate simulations of distortion and size distributions, combined with a pairwise training process, can close much of that gap. The pairwise step separates the optimization of general detection accuracy from robustness to those distortions. The approach placed third in the NTIRE Robust AI-Generated Image Detection in the Wild challenge. A reader would care because reliable in-the-wild detection is required to handle misinformation and synthetic media at scale.

Core claim

The authors show that a LoRA-based Pairwise Training strategy, built on targeted finetuning of a visual foundation model together with distortion and size simulations to match validation and test distributions, plus a pairwise process that decouples generalization from robustness optimization, delivers robust AIGI detection under severe distortions.

What carries the argument

LoRA-based Pairwise Training (LPT), which applies low-rank adaptation to a visual foundation model while using pairwise comparisons on simulated distorted pairs to separately tune generalization and robustness.

If this is right

  • Detectors maintain higher accuracy when images undergo compression, resizing, or other common alterations after generation.
  • Simulation of expected data distributions during training reduces the performance drop seen in wild deployment.
  • Pairwise training allows independent control over detection accuracy and distortion tolerance without one harming the other.
  • Low-rank adaptation makes it practical to adapt large visual models for this task without full retraining.

Where Pith is reading between the lines

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

  • The same decoupling idea could apply to other detection or classification tasks that face distribution shift after initial training.
  • More sophisticated or adaptive simulation of distortions might further narrow the remaining gap to real-world conditions.
  • The method's reliance on a specific foundation model raises the question of how performance scales when swapping to newer or smaller backbones.

Load-bearing premise

The simulated distortions and size variations introduced during training will sufficiently represent the unpredictable distortions that appear in actual deployment.

What would settle it

A new test set containing distortion types or combinations absent from the training simulations, on which the method shows no gain over standard fine-tuning of the same foundation model.

Figures

Figures reproduced from arXiv: 2604.12307 by Hao Sun, Qi Zhang, Ruiyang Xia, Xuelong Li, Yaowen Xu, Zhaofan Zou, Zhongjiang He.

Figure 1
Figure 1. Figure 1: Previous AIGI detection methods mainly consider [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distorted Images Illustration. Irrespective of whether the images are real or fake, all images are subjected to multiple distortions with high-level strength, such as Gaussian blur, quantization, impulse noise, brighteness, and tone curve transformations [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Image Area Statistics. There is a distribution mis￾matching between the training and validation/test sets. (b) Special Cases from Validation/Test sets. Some images only appear to have local semantic or non-semantic information. tic information, but the validation/test sets appear to have local semantic or non-semantic information (see [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The overall framework of LPT. To achieve robust detection, our strategy involves the simulation of data distribution, the fine￾tuning of the visual foundation model, and the pairwise training, respectively. Algorithm 1 Random Crop and Resize Augmentation Input: Input image img of shape (H, W, C), target size tgt, thresholds T1, T2 Output: Augmented image img 1: (H, W, C) ← img.shape 2: rc ← random sampling… view at source ↗
Figure 5
Figure 5. Figure 5: Left: The impact of the distortion strength. Right: The impact of additional data incorporation. 5.3.2. Different Sample Simulation To evaluate the effectiveness of the data distribution simu￾lation, we conduct ablation studies on both data scale varia￾tion and image distortion modeling. Tab. 5 shows the abla￾tion studies related to the image size and distortion. Consid￾ering the size variations in the val… view at source ↗
read the original abstract

The proliferation of highly realistic AI-Generated Image (AIGI) has necessitated the development of practical detection methods. While current AIGI detectors perform admirably on clean datasets, their detection performance frequently decreases when deployed "in the wild", where images are subjected to unpredictable, complex distortions. To resolve the critical vulnerability, we propose a novel LoRA-based Pairwise Training (LPT) strategy designed specifically to achieve robust detection for AIGI under severe distortions. The core of our strategy involves the targeted finetuning of a visual foundation model, the deliberate simulation of data distribution during the training phase, and a unique pairwise training process. Specifically, we introduce distortion and size simulations to better fit the distribution from the validation and test sets. Based on the strong visual representation capability of the visual foundation model, we finetune the model to achieve AIGI detection. The pairwise training is utilized to improve the detection via decoupling the generalization and robustness optimization. Experiments show that our approach secured the 3th placement in the NTIRE Robust AI-Generated Image Detection in the Wild challenge

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 proposes a LoRA-based Pairwise Training (LPT) strategy for robust AI-generated image (AIGI) detection under severe, unpredictable distortions. The method finetunes a visual foundation model via LoRA, introduces targeted distortion and size simulations during training to align with validation/test distributions, and applies a pairwise training process to separately optimize generalization and robustness. The central empirical claim is that this yields strong performance, evidenced by a third-place ranking in the NTIRE Robust AI-Generated Image Detection in the Wild challenge.

Significance. If the LPT components prove reproducible and the distortion simulations generalize beyond the specific challenge distribution, the work could provide a practical, parameter-efficient route to hardening AIGI detectors for real-world deployment. The combination of LoRA finetuning with explicit distribution simulation and pairwise decoupling is a targeted adaptation for the robustness setting. However, the significance is constrained by the absence of any reported quantitative metrics, ablation results, or out-of-distribution tests, leaving the contribution resting primarily on an unreported challenge ranking rather than transparent evidence.

major comments (3)
  1. [Abstract] Abstract: The claim that the method 'secured the 3th placement' is presented without any accompanying quantitative results (accuracy, AUC, or F1 on clean vs. distorted images), ablation studies on the LoRA, simulation, or pairwise components, or error bars. Because the central claim of robust detection rests on this empirical outcome, the lack of supporting numbers prevents evaluation of whether LPT actually drives the ranking.
  2. [Method] Method description (pairwise training): The text states that pairwise training 'decouples the generalization and robustness optimization,' yet supplies no details on pair construction, the specific loss terms, how positive/negative pairs are sampled, or the training schedule. This mechanism is load-bearing for the claimed improvement in robustness and must be specified for the contribution to be assessable.
  3. [Experiments] Experiments / simulation procedure: Distortion and size simulations are introduced 'to better fit the distribution from the validation and test sets,' but no analysis is given showing coverage of the distortion space or performance on novel combinations outside the challenge set. The robustness claim therefore risks being tied to the particular NTIRE test distribution rather than unpredictable real-world distortions.
minor comments (2)
  1. [Abstract] Abstract contains the typo '3th placement' (should be '3rd').
  2. [Abstract] The manuscript would be strengthened by including at least one summary table of key metrics (clean vs. distorted performance) early in the paper, even if full ablations appear later.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We address each major comment point by point below, indicating the revisions we will make to improve clarity, reproducibility, and evidential support.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim that the method 'secured the 3th placement' is presented without any accompanying quantitative results (accuracy, AUC, or F1 on clean vs. distorted images), ablation studies on the LoRA, simulation, or pairwise components, or error bars. Because the central claim of robust detection rests on this empirical outcome, the lack of supporting numbers prevents evaluation of whether LPT actually drives the ranking.

    Authors: We agree that the abstract, constrained by length, does not include supporting metrics or ablations. In the revised manuscript we will expand the abstract to report key quantitative results from the NTIRE challenge (accuracy, AUC on the test set) and will add a dedicated experiments subsection with ablation tables for the LoRA, distortion/size simulation, and pairwise components, including error bars from repeated runs where available. This will make the empirical contribution transparent and directly tied to the ranking. revision: yes

  2. Referee: [Method] Method description (pairwise training): The text states that pairwise training 'decouples the generalization and robustness optimization,' yet supplies no details on pair construction, the specific loss terms, how positive/negative pairs are sampled, or the training schedule. This mechanism is load-bearing for the claimed improvement in robustness and must be specified for the contribution to be assessable.

    Authors: We acknowledge that the current description of pairwise training is insufficient for reproducibility. In the revised version we will expand the method section with explicit details on pair construction (pairing each image with its simulated-distortion counterpart), the loss terms (standard cross-entropy for generalization and an additional robustness-oriented term), positive/negative pair sampling strategy, and the full training schedule with hyperparameters. This will clarify how the decoupling is achieved and allow independent assessment of the robustness gains. revision: yes

  3. Referee: [Experiments] Experiments / simulation procedure: Distortion and size simulations are introduced 'to better fit the distribution from the validation and test sets,' but no analysis is given showing coverage of the distortion space or performance on novel combinations outside the challenge set. The robustness claim therefore risks being tied to the particular NTIRE test distribution rather than unpredictable real-world distortions.

    Authors: We recognize the value of demonstrating that the simulations are not narrowly tuned to the challenge distribution. In the revision we will add a detailed breakdown of the distortion types and parameter ranges used, together with coverage analysis (e.g., tables or figures). We will also include performance results on a small number of novel distortion combinations outside the challenge set and a limitations paragraph discussing the extent of generalization to truly unpredictable real-world distortions. Because new large-scale OOD experiments would require additional compute, we mark this revision as partial. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical method validated externally

full rationale

The paper presents an empirical engineering approach (LoRA-based finetuning, deliberate distortion/size simulations, and pairwise training) whose central claim of robustness is supported by 3rd-place ranking on the external NTIRE challenge benchmark. No mathematical derivation, equations, or fitted parameters are presented as predictions that reduce to the inputs by construction. The simulations are described as a design choice to match challenge distributions, but this is not a self-referential reduction or load-bearing self-citation; success is measured against an independent test set rather than by internal consistency alone. The derivation chain is self-contained as a practical training procedure.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that simulated distortions during training will produce robustness to unseen real-world distortions; this is a standard domain assumption in robust learning but receives no independent verification in the abstract.

axioms (2)
  • domain assumption Visual foundation models possess strong representation capabilities that can be adapted via LoRA for AIGI detection
    Invoked when stating that finetuning the foundation model achieves detection.
  • domain assumption Pairwise training can decouple generalization and robustness optimization without harming overall performance
    Core justification for the pairwise component.

pith-pipeline@v0.9.0 · 5506 in / 1367 out tokens · 48081 ms · 2026-05-10T14:51:37.521541+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

45 extracted references · 14 canonical work pages · 4 internal anchors

  1. [1]

    Dual data alignment makes ai- generated image detector easier generalizable.arXiv preprint arXiv:2505.14359, 2025

    Ruoxin Chen, Junwei Xi, Zhiyuan Yan, Ke-Yue Zhang, Shuang Wu, Jingyi Xie, Xu Chen, Lei Xu, Isabel Guan, Taiping Yao, et al. Dual data alignment makes ai- generated image detector easier generalizable.arXiv preprint arXiv:2505.14359, 2025. 7

    work page arXiv 2025

  2. [2]

    Fair deepfake detec- tors can generalize

    Harry Cheng, Ming-Hui Liu, Yangyang Guo, Tianyi Wang, Liqiang Nie, and Mohan Kankanhalli. Fair deepfake detec- tors can generalize. InNeurIPS, 2025. 2

    2025

  3. [3]

    Meta clip 2: A worldwide scaling recipe.arXiv preprint arXiv:2507.22062,

    Yung-Sung Chuang, Yang Li, Dong Wang, Ching-Feng Yeh, Kehan Lyu, Ramya Raghavendra, James Glass, Lifei Huang, Jason Weston, Luke Zettlemoyer, et al. Meta clip 2: A world- wide scaling recipe.arXiv preprint arXiv:2507.22062, 2025. 7

    work page arXiv 2025

  4. [4]

    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. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023. 3

    2023

  5. [5]

    Forensics adapter: Adapting clip for generalizable face forgery detection

    Xinjie Cui, Yuezun Li, Ao Luo, Jiaran Zhou, and Junyu Dong. Forensics adapter: Adapting clip for generalizable face forgery detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2

    2025

  6. [6]

    Think twice before detect- ing GAN-generated fake images from their spectral domain imprints

    Chengyu Dong, Zhiqi Wang, Ting Yao, Jiale Liang, Shouhong Ding, Jiatong Li, et al. Think twice before detect- ing GAN-generated fake images from their spectral domain imprints. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    2022

  7. [7]

    Watch your up-convolution: Cnn based generative deep neural net- works are failing to reproduce spectral distributions

    Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural net- works are failing to reproduce spectral distributions. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2

    2020

  8. [8]

    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. InPro- ceedings of the 37th International Conference on Machine Learning (ICML), 2020. 2

    2020

  9. [9]

    Improving interpretabil- ity and robustness for the detection of ai-generated images

    Tatiana Gaintseva, Laida Kushnareva, German Magai, Irina Piontkovskaya, Sergey Nikolenko, Martin Benning, Serguei Barannikov, and Gregory Slabaugh. Improving interpretabil- ity and robustness for the detection of ai-generated images. arXiv preprint arXiv:2406.15035, 2024. 3

    work page arXiv 2024

  10. [10]

    NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild

    Aleksandr Gushchin, Khaled Abud, Ekaterina Shumitskaya, Artem Filippov, Georgii Bychkov, Sergey Lavrushkin, Mikhail Erofeev, Anastasia Antsiferova, Changsheng Chen, Shunquan Tan, Radu Timofte, Dmitriy Vatolin, et al. NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild . InProceedings of the IEEE/CVF Conference on Computer Vision and Pa...

    2026

  11. [11]

    Parameter-efficient transfer learning for nlp

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Maziar Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. InProceedings of the 36th Inter- national Conference on Machine Learning (ICML), 2019. 3

    2019

  12. [12]

    LoRA: Low-rank adaptation of large language mod- els

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language mod- els. InInternational Conference on Learning Representa- tions (ICLR), 2022. 3, 5

    2022

  13. [13]

    So-fake: Benchmarking and explain- ing social media image forgery detection.arXiv preprint arXiv:2505.18660, 2025

    Zhenglin Huang, Tianxiao Li, Xiangtai Li, Haiquan Wen, Yiwei He, Jiangning Zhang, Hao Fei, Xi Yang, Xiaowei Huang, Bei Peng, et al. So-fake: Benchmarking and explain- ing social media image forgery detection.arXiv preprint arXiv:2505.18660, 2025. 6

    work page arXiv 2025

  14. [14]

    Scaling up visual and vision-language representa- tion learning with noisy text supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yucen Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representa- tion learning with noisy text supervision. InProceedings of the 38th International Conference on Machine Learning (ICML), 2021. 2

    2021

  15. [15]

    Vi- sual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning. InEuropean Conference on Computer Vision (ECCV), 2022. 3 8

    2022

  16. [16]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4401–4410, 2019. 1

    2019

  17. [17]

    Navigating the challenges of ai- generated image detection in the wild: What truly matters? arXiv preprint arXiv:2507.10236, 2025

    Despina Konstantinidou, Dimitrios Karageorgiou, Chris- tos Koutlis, Olga Papadopoulou, Emmanouil Schinas, and Symeon Papadopoulos. Navigating the challenges of ai- generated image detection in the wild: What truly matters? arXiv preprint arXiv:2507.10236, 2025. 3

    work page arXiv 2025

  18. [18]

    Leveraging rep- resentations from intermediate encoder-blocks for synthetic image detection

    Christos Koutlis and Symeon Papadopoulos. Leveraging rep- resentations from intermediate encoder-blocks for synthetic image detection. InEuropean Conference on computer vi- sion, pages 394–411. Springer, 2024. 7

    2024

  19. [19]

    Learning real facial concepts for independent deepfake detection

    Ming-Hui Liu, Harry Cheng, Tianyi Wang, Xin Luo, and Xin-Shun Xu. Learning real facial concepts for independent deepfake detection. InIJCAI, pages 1585–1593. ijcai.org,

  20. [20]

    SGDR: Stochastic Gradient Descent with Warm Restarts

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochas- tic gradient descent with warm restarts.arXiv preprint arXiv:1608.03983, 2016. 6

    work page internal anchor Pith review arXiv 2016

  21. [21]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 6

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  22. [22]

    Detecting gan-generated images by orthogo- nal training of multiple cnns

    Sara Mandelli, Nicol `o Bonettini, Paolo Bestagini, and Ste- fano Tubaro. Detecting gan-generated images by orthogo- nal training of multiple cnns. In2022 IEEE International Conference on Image Processing (ICIP), pages 3091–3095. IEEE, 2022. 3

    2022

  23. [23]

    Do GANs leave artificial fingerprints? arXiv preprint arXiv:1812.11842, 2019

    Francesco Marra, Diego Gragnaniello, Davide Cozzolino, and Luisa Verdoliva. Do GANs leave artificial fingerprints? arXiv preprint arXiv:1812.11842, 2019. 2

    work page arXiv 2019

  24. [24]

    Towards uni- versal fake image detectors that generalize across generative models

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards uni- versal fake image detectors that generalize across generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2

    2023

  25. [25]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 4195–4205,

  26. [26]

    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. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 2, 6, 7

    2021

  27. [27]

    Im- proving robustness against common corruptions with fre- quency biased models

    Tonmoy Saikia, Cordelia Schmid, and Thomas Brox. Im- proving robustness against common corruptions with fre- quency biased models. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 10211– 10220, 2021. 5

    2021

  28. [28]

    Shield: An evaluation benchmark for face spoofing and forgery detection with multimodal large lan- guage models.Visual Intelligence, 3(1):9, 2025

    Yichen Shi, Yuhao Gao, Yingxin Lai, Hongyang Wang, Jun Feng, Lei He, Jun Wan, Changsheng Chen, Zitong Yu, and Xiaochun Cao. Shield: An evaluation benchmark for face spoofing and forgery detection with multimodal large lan- guage models.Visual Intelligence, 3(1):9, 2025. 1

    2025

  29. [29]

    Flava: A foundational language and vision alignment model

    Amanpreet Singh et al. Flava: A foundational language and vision alignment model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    2022

  30. [30]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  31. [31]

    Towards general visual-linguistic face forgery detection

    Kaixin Sun et al. Towards general visual-linguistic face forgery detection. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR),

  32. [32]

    EVA-CLIP: Improved Training Techniques for CLIP at Scale

    Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 5, 7

    work page internal anchor Pith review arXiv 2023

  33. [33]

    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. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 28130–28139, 2024. 7

    2024

  34. [34]

    Seman- tic visual anomaly detection and reasoning in ai-generated images.arXiv preprint arXiv:2510.10231, 2025

    Chuangchuang Tan, Xiang Ming, Jinglu Wang, Renshuai Tao, Bin Li, Yunchao Wei, Yao Zhao, and Yan Lu. Seman- tic visual anomaly detection and reasoning in ai-generated images.arXiv preprint arXiv:2510.10231, 2025. 2

    work page arXiv 2025

  35. [35]

    Oddn: Addressing unpaired data challenges in open-world deepfake detection on online social networks

    Renshuai Tao, Manyi Le, Chuangchuang Tan, Huan Liu, Haotong Qin, and Yao Zhao. Oddn: Addressing unpaired data challenges in open-world deepfake detection on online social networks. InProceedings of the AAAI Conference on Artificial Intelligence, pages 799–807, 2025. 1

    2025

  36. [36]

    Improving robustness to corruptions with multiplica- tive weight perturbations.Advances in Neural Information Processing Systems, 37:35492–35516, 2024

    Trung Trinh, Markus Heinonen, Luigi Acerbi, and Samuel Kaski. Improving robustness to corruptions with multiplica- tive weight perturbations.Advances in Neural Information Processing Systems, 37:35492–35516, 2024. 5

    2024

  37. [37]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695–8704, 2020. 1, 2, 3, 7, 8

    2020

  38. [38]

    A sanity check for ai-generated image detection.arXiv preprint arXiv:2406.19435, 2024

    Shilin Yan, Ouxiang Li, Jiayin Cai, Yanbin Hao, Xi- aolong Jiang, Yao Hu, and Weidi Xie. A sanity check for ai-generated image detection.arXiv preprint arXiv:2406.19435, 2024. 6

    work page arXiv 2024

  39. [39]

    Orthogonal subspace decompo- sition for generalizable ai-generated image detection.arXiv preprint arXiv:2411.15633, 2024

    Zhiyuan Yan, Jiangming Wang, Peng Jin, Ke-Yue Zhang, Chengchun Liu, Shen Chen, Taiping Yao, Shouhong Ding, Baoyuan Wu, and Li Yuan. Orthogonal subspace decompo- sition for generalizable ai-generated image detection.arXiv preprint arXiv:2411.15633, 2024. 7

    work page arXiv 2024

  40. [40]

    Attributing fake im- ages to GANs: Learning and analyzing GAN fingerprints

    Ning Yu, Larry Davis, and Mario Fritz. Attributing fake im- ages to GANs: Learning and analyzing GAN fingerprints. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2

    2019

  41. [41]

    Mlep: Multi-granularity local entropy patterns for generalized ai-generated image detection

    Lin Yuan, Xiaowan Li, Yan Zhang, Jiawei Zhang, Hongbo Li, and Xinbo Gao. Mlep: Multi-granularity local entropy patterns for generalized ai-generated image detection. InThe Thirty-ninth Annual Conference on Neural Information Pro- cessing Systems, 2025. 7

    2025

  42. [42]

    Lit: Zero-shot transfer with locked-image text tuning

    Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2022. 2 9

    2022

  43. [43]

    Adalora: Adap- tive budget allocation for parameter-efficient fine-tuning

    Qingru Zhang, Minshuo Chen, Anton Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: Adap- tive budget allocation for parameter-efficient fine-tuning. InInternational Conference on Learning Representations (ICLR), 2023. 3

    2023

  44. [44]

    Detect- ing and simulating artifacts in GAN fake images.arXiv preprint arXiv:1907.06515, 2019

    Xu Zhang, Sidharth Karaman, and Shih-Fu Chang. Detect- ing and simulating artifacts in GAN fake images.arXiv preprint arXiv:1907.06515, 2019. 2

    work page arXiv 1907

  45. [45]

    Breaking semantic artifacts for gener- alized ai-generated image detection

    Cheng Zheng et al. Breaking semantic artifacts for gener- alized ai-generated image detection. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 2 10

    2024