Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Few-Shot Synthetic Image Attribution: Identifying Unseen Generators with Limited Samples

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that AI-generated images can be traced to their source model even when that model was never seen in training, using only a handful of reference images, and that a unified detector–attributor trained on a 45-generator datas

desk verdict Useful dataset and a sensible task framing, but the central few-shot attribution claim is only tested on held-out OmniFake categories, so the open-set generalization story is not yet supported. read the letter →

arxiv 2509.25682 v2 pith:LO7EFIOW submitted 2025-09-30 cs.CV

Shiyu Wu , Shuyan Li , Jing Li , Jing Liu , Yequan Wang This is my paper
classification cs.CV
keywords AI-generatedimagedetectionfew-shotattributionopen-setrecognitionsourceidentificationcontrastivelearningsyntheticdatasetdeepfakeforensics
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

The paper's central claim is that source attribution of synthetic images can be pushed beyond closed-set recognition into an open-set, few-shot setting: given just five reference images from a previously unseen generator, a model can both say an image is fake and name the generator that made it. To make this work, the authors built OmniFake, a dataset of 1.17 million images from 45 architecturally distinct generators spanning GANs, diffusion, autoregressive, and hybrid models. On this dataset, their OmniDFA framework reports 72.86% 5-way and 50.24% 15-way few-shot attribution on held-out categories, and top zero-shot detection accuracy on external benchmarks (95.86 on GenImage, 83.48 on Chameleon). A sympathetic reader would take away that explicit, fine-grained modeling of generator-specific artifacts—rather than a binary fake/real split—is a route to detectors that do not go stale as new generators appear.

What carries the argument

The load-bearing mechanism is a contrastively learned hypersphere embedding: L2-normalized feature vectors are trained with supervised contrastive loss across the 45 generator classes (real images held as one class), so known forgery types occupy separated regions and unseen generators, sharing partial artifacts with known ones, tend to land between them. A sphere center loss with a learnable, momentum-updated boundary threshold seals real images into a compact region, and a dual-path ConvNeXt encoder (global aspect-ratio-preserving resize plus high-resolution local crops, fused by an MLP) supplies the features. Few-shot attribution is then prototypical: mean support-set embeddings of an uns

What would settle it

Take a recent generator from a family absent from the 45 (for example a streaming diffusion-transformer or a video-diffusion model), extract five support images, and run open-set attribution; if accuracy at identifying that source drops to chance while detection still works, the between-known-regions assumption fails.

Watch

Extended reading notes

Core claim

OmniDFA's central discovery is that model-specific biases are sparse, learnable, and quantifiable: a supervised contrastive loss can carve a feature space where each known generator occupies a distinct region, while a center-loss term pins real images into a compact cluster bounded by a learned angular threshold. At test time, images are embedded with a dual-path extractor that combines a low-resolution global view with high-resolution local crops; a 5-shot support set from an unseen generator suffices to build a prototype, and query images are attributed by similarity to those prototypes. The paper reports 72.86% 5-way and 50.24% 15-way open-set attribution on held-out OmniFake categories a

Load-bearing premise

Unseen generators' images will fall between the feature regions of known generators, sharing partial artifacts; if a genuinely novel architecture produces features that lie far outside the learned manifold, five-shot prototypes cannot locate it.

Editorial extensions

If this is right

  • A single trained model can serve as both authenticity detector and source attributor, removing the need to retrain a network whenever a new generation model is released.
  • Adding a newly released generator to the system costs only a few labeled images, not a full dataset and retraining run.
  • If the 72.86%/50.24% open-set few-shot figures on held-out categories hold, then fine-grained generator artifacts are stable enough to support forensic tracing in practice.
  • Training on many architecturally distinct generators appears to transfer: zero-shot accuracy on GenImage and Chameleon beats prior detectors, suggesting attribution-style supervision is a viable generalization strategy.

Reading between the lines

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

  • Editorial inference: the same embedding space should support unsupervised discovery of novel generator families by clustering unknown images, a capability the paper does not demonstrate.
  • Editorial inference: the reported results are for exactly five support images; a natural extension is measuring how attribution accuracy degrades as support sets shrink to one image or grow to ten, which would tell practitioners how many samples they really need.
  • Editorial inference: the dataset's insistence on architecturally distinct generators is probably a necessary condition for the few-shot transfer—datasets full of fine-tuned versions of one backbone would likely not reproduce these numbers.
  • Editorial inference: the method should be testable on video and audio generators; if model-specific traces persist across modalities, the same contrastive-prototype recipe could generalize beyond still images.
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 / 4 minor

Summary. The paper proposes a new paradigm, open-set few-shot synthetic image attribution: given a few reference images from a generator not seen during training, the model should identify the source. The authors construct OmniFake, a 45-generator, 1.17M-image dataset with architecturally distinct fake categories, and introduce OmniDFA, a dual-path global/local framework trained with a supervised contrastive loss on fake categories, a sphere center loss on real images, and a Tukey-fence authenticity boundary. Experiments report state-of-the-art within-OmniFake few-shot attribution (72.86% 5-way, 50.24% 15-way) and zero-shot detection on OmniFake, GenImage, and Chameleon, together with ablations and robustness analyses.

Significance. If validated, the paper would make a useful contribution: OmniFake is substantially broader in generator diversity than existing attribution datasets, the task formulation is practically motivated, and the proposed method is simple and reproducible. The authors also use official weights for most baselines, run 10,000 episodes for the few-shot protocol, and promise code and dataset release. The main limitation is that the central generalization claim is not yet established: few-shot attribution is evaluated only on held-out OmniFake categories sampled from the same curation pipeline, and the cross-dataset experiments exercise detection only, not attribution. The reported attribution gains over FSD are also small and lack variance estimates. These gaps are addressable and should be closed before publication.

major comments (3)
  1. [Section 3.2 / Section 4.3 / Table 2] The central claim that OmniDFA 'reliably identifies unseen generators' is supported only by 3-fold cross-validation on OmniFake, where the held-out categories are other generators from the same 45-model corpus, collected through the same pipeline (same prompt sources, resolutions, and benchmark origins). They are unseen as labels, not as a distribution shift. The method depends on the assertion in Section 3.2 that 'unseen fake samples are likely to fall between these regions', but the paper provides no geometric validation of this interpolation assumption, and no attribution experiment against a genuinely external generator is reported. Table 4 is binary detection only, and Chameleon lacks generator labels. If a novel architecture maps outside the span of known feature regions, nearest-centroid attribution has no basis. Please add an external source-labeled attribution test (e.g., hold o
  2. [Tables 2-4] No measure of variance is reported anywhere. After 10,000 episodes, Table 2 shows OmniDFA at 72.86% vs FSD 71.14% (5-way) and 50.24% vs 49.71% (15-way); without standard deviations or confidence intervals these differences may not be statistically significant. Table 3 and Table 4 also report single-run numbers. Report mean and standard deviation over episodes and/or training seeds, and use paired tests across identical episodes when comparing against fixed checkpoints.
  3. [Section 4.5 / Table 4] The 'strictly zero-shot' claim for GenImage needs a leakage audit. OmniFake is constructed in part from GenImage and WildFake images (Section 2.1), and Table 5 lists ADM, GLIDE, Midjourney V5, SD1.5, VQDM, and other GenImage/WildFake generators in the training portion. The text says only that SD1.5, SDXL, and SD3-Medium were explicitly removed before the cross-dataset run. If any of the other GenImage categories remain in training, the GenImage column is not truly zero-shot. For Chameleon, which has no fine-grained category labels, it is unclear how all 'related model families' could be excluded. Please provide a precise list of removed and remaining training generators for each cross-dataset run, and check for near-duplicate images between OmniFake training and the test benchmarks.
minor comments (4)
  1. [Throughout] Typos and inconsistencies: 'OminFake' (Section 4.4), 'Chemeleon' (Section 4.5), 'Shpere' (Figure 3 caption), 'V AEs' (Section 2.1), and inconsistent spelling of 'DiT-XL/2' vs 'DIT-XL/2' in Table 5.
  2. [Section 4.5 / Appendix E.2] The cross-dataset experiments use 40 training epochs and lambda = 5e-3, while the main experiments use 20 epochs and lambda = 0.01. Please state whether these hyperparameters were selected on a validation split disjoint from the GenImage/Chameleon test sets, or were fixed before seeing test results.
  3. [Figure 5] The ablation bar chart would benefit from exact numeric values, error bars, and significance tests; the current figure is hard to read and differences between components are not quantified.
  4. [Equation (4)] Please clarify the data over which Q1 and Q3 are computed (per mini-batch, per epoch, or a running set), how the initial boundary gamma is set, and how the momentum update interacts with the Tukey-fence estimate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: OmniDFA's few-shot attribution uses standard supervised contrastive embeddings plus prototype classification; the 'unseen generators fall between known regions' statement is an empirical generalization assumption, not an equation-level reduction. The self-cited FSD is used only as a baseline and is not load-bearing.

full rationale

The paper's derivation chain does not reduce to its inputs. The few-shot attribution mechanism is a standard pipeline: supervised contrastive loss (Eq. 1) learns known-class embeddings, sphere center loss (Eq. 2) clusters real features, and open-set classification is performed by nearest-prototype matching on a support set. No fitted parameter is renamed as a prediction, and no self-defined quantity is derived from the target result. The main potential concern is Section 3.2's assumption that "unseen fake samples are likely to fall between these regions, as they share partial artifacts with known forgery types." This is an empirical claim about feature-space geometry, not a circular derivation; if false for genuinely external generators, the method would fail, but that is a generalization/external-validity risk, not circularity. The few-shot evaluation (Table 2) uses 3-fold cross-validation on OmniFake, so the 'unseen' generators are held-out categories from the same curated corpus; this limits how strongly the open-set claim is supported, but the evaluation protocol is standard and the held-out classes are not used during training. Cross-dataset GenImage/Chameleon tests exercise binary detection only, not attribution, but again this is an evaluation-coverage concern. The self-citation to FSD (Wu et al., 2025b) appears as a baseline and a related-work comparison; it is not invoked to justify the central derivation or to forbid alternatives, so it is not load-bearing. No circular step can be exhibited by quoting equations that reduce to one another or by showing a fitted parameter masquerading as a prediction. Therefore the score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The framework relies on standard network components and training losses. The free parameters are hyperparameters and a data-fitted decision boundary. The core assumptions are about feature transferability and the representativeness of the dataset.

free parameters (5)
  • temperature tau = 0.07
    Temperature in supervised contrastive loss (Eq. 1), chosen by hand.
  • center loss weight lambda = 0.01
    Scaling factor in final objective (Eq. 3), chosen by hand.
  • momentum coefficient beta = 0.99
    Momentum for boundary threshold update (Eq. 4), chosen by hand.
  • Tukey fence multiplier = 1.5
    IQR multiplier in boundary threshold calculation (Eq. 4), standard rule but hand-chosen.
  • boundary threshold gamma = updated via momentum, not fixed
    Learned from real sample deviations via Tukey's fences; effectively a fitted decision boundary.
assumptions (4)
  • domain assumption Pretrained ImageNet features transfer to synthetic image forensics
    Uses ConvNeXt-Small pretrained on ImageNet as feature extractor; assumes these features capture forensically relevant artifacts.
  • domain assumption The feature manifold is such that unseen generators' features lie between known generators' features
    Sec 3.2: 'unseen fake samples are likely to fall between these regions, as they share partial artifacts with known forgery types.' This is load-bearing for few-shot attribution.
  • domain assumption OmniFake's 45 generators are representative of the space of generative models
    The zero-shot generalization claims rest on the diversity of OmniFake; if it underrepresents future architectures, few-shot attribution may fail.
  • domain assumption The 3-fold split creates truly unseen generators with no residual artifact leakage from training categories
    They ensure distinct architectures, but some may share backbones (e.g., SD variants) which could leak discriminative features; no leakage analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Few-Shot Synthetic Image Attribution: Identifying Unseen Generators with Limited Samples." pith.science (2026). https://pith.science/paper/LO7EFIOW

@misc{pith2026250925682,
  author       = {Pith},
  title        = {Pith review of: Few-Shot Synthetic Image Attribution: Identifying Unseen Generators with Limited Samples},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LO7EFIOW}},
  note         = {Machine review of arXiv:2509.25682}
}
abstract

AI-generated image (AIGI) attribution presents a pressing challenge that goes beyond mere AIGI detection, aiming to identify the source model or technique responsible for a synthetic image. However, most previous source attribution methods operate in a closed-set manner, which necessitates retraining to recognize any novel category, preventing adaptation to the rapid evolution of image generation. In this work, we propose a new paradigm for synthetic image attribution, termed few-shot attribution. This paradigm targets the reliable identification of unseen generators using only limited samples, making it highly suitable for real-world applications. To facilitate this work, we construct OmniFake, a large-scale, well-categorized synthetic image dataset that contains $1.17$ million images from $45$ distinct generators. We further introduce OmniDFA (Omni Detector and Few-shot Attributor), a few-shot attribution baseline that not only assesses the authenticity of images but also determines their synthesis origins. Experiments demonstrate that OmniDFA exhibits excellent capability in few-shot attribution and achieves state-of-the-art generalization performance in AIGI detection. Our dataset and code are available at https://github.com/teheperinko541/OmniDFA.

Figures

Figures reproduced from arXiv: 2509.25682 by the authors.

Figure 1
Figure 1. Comparison of task-specific pipelines for synthetic analysis. Our new attribution paradigm [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Samples of the OmniFake dataset. Our dataset covers a broad spectrum of generative [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of OmniDFA. Our dual-path architecture captures both low-level and high-level [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Robustness to compression and blurring. The vertical lines indicate the perturbation [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Ablation study on model component. we found that the higher the degree of Gaussian blur, the more likely an image is to be perceived as synthetic, which is contrary to the common assumption that blurring makes real and synthetic images harder to distinguish. This provi…

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. Full citation record

  1. LaP-Forensics: Latent-Pixel Consistency Guided Multimodal Reasoning for Deepfake Detection

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A dual-stream deepfake forensic model that adds DDIM reconstruction residuals to RGB features improves artifact localization and cross-generator detection in evaluations, with honest caveats about text faithfulness.

Reference graph

Works this paper leans on

54 extracted references · 14 linked inside Pith · cited by 1 Pith paper

  1. [4]

    Con- trasting deepfakes diffusion via contrastive learning and global-local similarities

    Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, Alessandro Nicolosi, and Rita Cucchiara. Con- trasting deepfakes diffusion via contrastive learning and global-local similarities. InComputer Vision - ECCV 2024 - 18th European Conference, volume 15121 ofLecture Notes in Computer Science, pp. 199–216. Springer,

    2024

  2. [5]

    Raising the bar of ai-generated image detection with CLIP

    Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with CLIP. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 - Workshops, pp. 4356–4366. IEEE, 2024a. Davide Cozzolino, Giovanni Poggi, Matthias Nießner, and Luisa Verdoliva. Zero-shot detection...

    2024

  3. [6]

    URLhttps://arxiv.org/abs/2505.14683. David C. Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. InIEEE/CVF International Conference on Computer Vision, ICCV 2023 - Workshops, Paris, France, October 2-6, 2023, pp. 382–392. IEEE, 2023a. David C. Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection...

    arXiv 2023

  4. [7]

    Shengbang Fang, Tai D

    URLhttps://arxiv.org/abs/ 2403.03206. Shengbang Fang, Tai D. Nguyen, and Matthew C. Stamm. Open set synthetic image source attri- bution. In34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023, pp

    arXiv 2023

  5. [9]

    URLhttps: //arxiv.org/abs/2503.21210. Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. InAnnual Conference on Neural Information Processing Systems 2014, pp. 2672–2680,

    arXiv 2014

  6. [11]

    Wukong: A 100 million large- scale chinese cross-modal pre-training benchmark

    Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Niu Minzhe, Xiaodan Liang, Lewei Yao, Runhui Huang, Wei Zhang, Xin Jiang, Chunjing Xu, and Hang Xu. Wukong: A 100 million large- scale chinese cross-modal pre-training benchmark. InAnnual Conference on Neural Information Processing Systems 2022, NeurIPS 2022,

    2022

  7. [12]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InAnnual Conference on Neural Information Processing Systems 2020, NeurIPS 2020,

    2020

  8. [13]

    Progressive growing of gans for im- proved quality, stability, and variation

    11 Under review as a conference paper at ICLR 2026 Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for im- proved quality, stability, and variation. In6th International Conference on Learning Representa- tions, ICLR

    2026

Show all 54 references
  1. [16]

    Ziqiang Li, Jiazhen Yan, Ziwen He, Kai Zeng, Weiwei Jiang, Lizhi Xiong, and Zhangjie Fu

    URLhttps://arxiv.org/abs/2405.08748. Ziqiang Li, Jiazhen Yan, Ziwen He, Kai Zeng, Weiwei Jiang, Lizhi Xiong, and Zhangjie Fu. Is artificial intelligence generated image detection a solved problem?, 2025b. URLhttps:// arxiv.org/abs/2505.12335. Ziyou Liang, Weifeng Liu, Run Wang...

    arXiv

  2. [17]

    Tsung-Yi Lin, Michael Maire, Serge J

    URLhttps://arxiv.org/ abs/2506.03147. Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. InComputer Vision - ECCV 2014 - 13th European Conference, volume 86...

    2014 arXiv

  3. [18]

    Which model generated this image? A model-agnostic approach for origin attribution

    Fengyuan Liu, Haochen Luo, Yiming Li, Philip Torr, and Jindong Gu. Which model generated this image? A model-agnostic approach for origin attribution. InComputer Vision - ECCV 2024 - 18th European Conference, volume 15120 ofLecture Notes in Computer Science, pp. 282–301. Springer,

    2024

  4. [19]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, pp. 11966–11976. IEEE,

    2022

  5. [20]

    Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding

    URLhttps://arxiv.org/abs/2304.13023. Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare 2: Latent reconstruction error based method for diffusion-generated image detection. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, pp. 17006–17015. IEEE,

    2024 arXiv

  6. [22]

    Community forensics: Using thousands of generators to train fake image detectors

    Jeongsoo Park and Andrew Owens. Community forensics: Using thousands of generators to train fake image detectors. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, pp. 8245–8257. Computer Vision Foundation / IEEE,

    2025

  7. [23]

    Anirudh Sundara Rajan and Yong Jae Lee

    URLhttps://arxiv.org/abs/ 2302.11970. Anirudh Sundara Rajan and Yong Jae Lee. Stay-positive: A case for ignoring real image features in fake image detection,

    arXiv

  8. [24]

    Jonas Ricker, Denis Lukovnikov, and Asja Fischer

    URLhttps://arxiv.org/abs/2502.07778. Jonas Ricker, Denis Lukovnikov, and Asja Fischer. AEROBLADE: training-free detection of latent diffusion images using autoencoder reconstruction error. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, pp. 9130–91...

    2024 arXiv

  9. [25]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aber- man. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, pp. 22500– 22510. IEEE,

    2023

  10. [26]

    URLhttps://arxiv.org/abs/2111. 02114. Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. DE-FAKE: detection and attribution of fake images generated by text-to-image generation models. InProceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, CCS 202...

    2023

  11. [27]

    Contrastive pseudo learning for open-world deepfake attribution

    Zhimin Sun, Shen Chen, Taiping Yao, Bangjie Yin, Ran Yi, Shouhong Ding, and Lizhuang Ma. Contrastive pseudo learning for open-world deepfake attribution. InIEEE/CVF International Conference on Computer Vision, ICCV 2023, pp. 20825–20835. IEEE,

    2023

  12. [28]

    Rethinking the up-sampling operations in cnn-based generative network for generalizable deep- fake detection

    Chuangchuang Tan, Huan Liu, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deep- fake detection. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, ...

    2024

  13. [29]

    Ovis-u1 technical report, 2025a

    Guo-Hua Wang, Shanshan Zhao, Xinjie Zhang, Liangfu Cao, Pengxin Zhan, Lunhao Duan, Shiyin Lu, Minghao Fu, Xiaohao Chen, Jianshan Zhao, Yang Li, and Qing-Guo Chen. Ovis-u1 technical report, 2025a. URLhttps://arxiv.org/abs/2506.23044. 13 Under review as a conference paper at ICL...

    2026 arXiv

  14. [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. InIEEE/CVF International Confer- ence on Computer Vision, ICCV 2023, pp. 22388–22398. IEEE, 2023a. Zhenting Wang, Chen Chen, Yi Zeng, Lin...

    2023

  15. [32]

    Detecting origin attribution for text-to-image diffu- sion models

    Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Detecting origin attribution for text-to-image diffu- sion models. InIEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2025, Tucson, AZ, USA, February 26 - March 6, 2025, pp. 8775–8785. IEEE,

    2025

  16. [33]

    Deepfake network architecture attribution

    Tianyun Yang, Ziyao Huang, Juan Cao, Lei Li, and Xirong Li. Deepfake network architecture attribution. InThirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, pp. 4662–

    2022

  17. [34]

    Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, and Xinpeng Zhang

    URLhttps://arxiv.org/abs/1506.03365. Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Patchcraft: Exploring texture patch for efficient ai-generated image detection,

    arXiv

  18. [35]

    Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang

    URLhttps://arxiv.org/abs/ 2311.12397. Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. InAnnual Conference on Neural Information Proces...

    2023 arXiv

  19. [36]

    For instance, the widely-used GenImage dataset (Zhu et al.,

    14 Under review as a conference paper at ICLR 2026 A NECESSITY OFCLASS-AWAREDATASET Most currently available datasets fail to account for the architectural uniqueness of generative mod- els. For instance, the widely-used GenImage dataset (Zhu et al.,

    2026

  20. [37]

    This structural homogeneity results in remarkably similar feature distributions among their gener- ated images

    includes three models: Stable Diffusion 1.4, Stable Diffusion 1.5, and Wukong, all of which share identical backbone structures. This structural homogeneity results in remarkably similar feature distributions among their gener- ated images. Such design leads to significant lim...

    2023

  21. [38]

    Therefore, merely making minor modifications to the weights or model architecture is insufficient to enable the model to express fundamentally different representations

    and LoRA adaptation (Hu et al., 2022). Therefore, merely making minor modifications to the weights or model architecture is insufficient to enable the model to express fundamentally different representations. To comprehensively evaluate the generalization ability of deepfake d...

    2022

  22. [39]

    As diffusion mod- els (Ho et al.,

    and LSUN (Yu et al., 2016), then tested across various other models. As diffusion mod- els (Ho et al.,

    2016

  23. [40]

    However, these datasets only cover a limited number of generators, restricting the generalization capability of detectors

    introduces the first million-scale synthetic dataset, paired with real images from ImageNet (Russakovsky et al., 2015). However, these datasets only cover a limited number of generators, restricting the generalization capability of detectors. Recent studies (Asnani et al.,

    2015

  24. [41]

    Community Forensics (Park & Owens,

    or LoRA adaption (Hu et al., 2022). Community Forensics (Park & Owens,

    2022

  25. [42]

    addresses the diversity limitation by aggregating samples from thousands of genera- tive models. However, these datasets are suboptimal for attribution, as their categories often consist of fine-tuned or adapted versions of the same base model, making it difficult to achieve m...

    2025

  26. [43]

    Recent studies (Cocchi et al.,

    utilizes 15 Under review as a conference paper at ICLR 2026 vision-language models which not only effectively incorporate textual prompts but also produce human-interpretable feature descriptions. Recent studies (Cocchi et al.,

    2026

  27. [44]

    Bi et al

    have introduced more sophisticated learning objectives to over- come the limitations of binary classifiers. Bi et al. (2023) introduce learning with real images only by analyzing pixel-level distributions. NTF (Liang et al.,

    2023

  28. [45]

    Inspired by these advances, we aim to fully explore the potential of metric learning for out-of-distribution deepfake detection

    employs self-supervised feature mapping to enhance transfer learning, while FSD (Wu et al., 2025b) learns a specialized metric space to distinguish unseen fake images with given samples. Inspired by these advances, we aim to fully explore the potential of metric learning for o...

    2020

  29. [46]

    Their motivations are based on the observation that models with different architectures exhibit distinct fingerprints

    focus on closed-set classification over GANs, where all generators encoun- tered during testing are assumed to be known at training time. Their motivations are based on the observation that models with different architectures exhibit distinct fingerprints. However, the closed-...

    2023

  30. [47]

    employs the CLIP model to attribute fake images created by text-to image diffusion models. Nevertheless, most methods still struggle to fully exploit fine-grained cross-modal fingerprints and show limited generalization capability in zero-shot scenarios when confronted with no...

    2023

  31. [48]

    Our synthetic categories include models from the same family, but we rigorously ensure that they are not derived from the same backbone

    Our OmniFake contains1.17M synthetic images generated by 45 distinct generative models, span- ning a broad spectrum of architectures. Our synthetic categories include models from the same family, but we rigorously ensure that they are not derived from the same backbone. To ens...

    2026

  32. [49]

    for in-the-wild descriptions. To maintain representativeness and diversity in prompt selection, we balance the sampling ratio among these sources at2 : 5 : 3, effectively integrating their respective characteristics. Based on these curated prompt lists, we lever- age17distinct...

    2026

  33. [50]

    However, this practice can introduce significant real biases by limiting the diversity and representativeness of the authentic class

    often rely on a single source dataset to represent authentic images. However, this practice can introduce significant real biases by limiting the diversity and representativeness of the authentic class. To mitigate this limitation and establish a more robust and generalizable ...

    2025

  34. [51]

    Table 6: Collection of datasets for real images in our OmniFake dataset. Real Image Dataset Training Test Laion-5B 251k 20k Wukong 242k 20k ImageNet-1k 174k 15k CC12M 160k 15k MSCOCO 113k 10k FFHQ 68k 2k CelebA-HQ 28k 2k LSUN-church 80k 2k IMD2020 33k 2k FODB 21k 2k Our authen...

    2021

  35. [52]

    These are balanced with carefully constructed datasets including Ima- geNet (Russakovsky et al., 2015), MSCOCO (Lin et al.,

    provide essential in-the-wild diver- sity by capturing the unfiltered heterogeneity of real-world internet imagery at scale, which is crucial for evaluating forensic models under realistic conditions. These are balanced with carefully constructed datasets including Ima- geNet ...

    2015

  36. [53]

    This approach not only maximizes the utilization of limited data but also guarantees that model performance is evaluated across diverse and representative subsets

    In each validation round, two folds are used for training, while the remaining fold serves as the held-out test set. This approach not only maximizes the utilization of limited data but also guarantees that model performance is evaluated across diverse and representative subse...

    2026

  37. [54]

    In Table 8, p denotes the probability of applying the corresponding transformation. It should be noted that in RandAugment (Cubuk et al., 2020), we deliberately exclude shear and translate transformations to ensure that our local feature extractor does not capture artifacts fr...

    2020

  38. [2014]

    Fake or jpeg? revealing common biases in generated image detection datasets

    Patrick Grommelt, Louis Weiss, Franz-Josef Pfreundt, and Janis Keuper. Fake or jpeg? revealing common biases in generated image detection datasets. InComputer Vision - ECCV 2024 Work- shops, volume 15644 ofLecture Notes in Computer Science, pp. 80–95. Springer,

    2024

  39. [2016]

    Omnigen2: Exploration to advanced multimodal generation, 2025a

    Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, Ze Liu, Ziyi Xia, Chaofan Li, Haoge Deng, Jiahao Wang, Kun Luo, Bo Zhang, Defu Lian, Xinlong Wang, Zhongyuan Wang, Tiejun Huang, and Zheng Liu. Omnigen2: Ex...

    arXiv

  40. [2018]

    Supervised contrastive learning

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. InAnnual Conference on Neural Information Processing Systems 2020, NeurIPS 2020,

    2020

  41. [2020]

    Improving synthetic image detection towards generalization: An image transformation perspective

    Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, and Fuli Feng. Improving synthetic image detection towards generalization: An image transformation perspective. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, V .1, KDD 2025, p...

    2025

  42. [2021]

    Towards universal fake image detectors that generalize across generative models

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, pp. 24480–24489. IEEE, 2023a. 12 Under review as a conference paper at ICLR ...

    2023

  43. [2022]

    Conceptual 12m: Pushing web- scale image-text pre-training to recognize long-tail visual concepts

    Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web- scale image-text pre-training to recognize long-tail visual concepts. InIEEE Conference on Com- puter Vision and Pattern Recognition, CVPR 2021, pp. 3558–3568. Computer Vision Foundatio...

    2021

  44. [2023]

    Leveraging frequency analysis for deep fake image recognition

    Joel Frank, Thorsten Eisenhofer, Lea Sch ¨onherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. InProceedings of the 37th International Conference on Machine Learning, ICML 2020, volume 119 ofProceedings of Ma...

    2020

  45. [2024]

    Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset, 2025a

    Jiuhai Chen, Zhiyang Xu, Xichen Pan, Yushi Hu, Can Qin, Tom Goldstein, Lifu Huang, Tianyi Zhou, Saining Xie, Silvio Savarese, Le Xue, Caiming Xiong, and Ran Xu. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset, 2025a. URLhttps: //arx...

    2024 arXiv

  46. [2025]

    Tu Bui, Ning Yu, and John P

    URLhttps://arxiv.org/abs/2407.20020. Tu Bui, Ning Yu, and John P. Collomosse. Repmix: Representation mixing for robust attribution of synthesized images. InComputer Vision - ECCV 2022, volume 13674 ofLecture Notes in Computer Science, pp. 146–163. Springer,

    2022 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.