Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

Bridging the Gap Between Ideal and Real-world Evaluation: Benchmarking AI-Generated Image Detection in Challenging Scenarios

T0 review · 3 major / 6 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 current AI-generated image detectors lose large amounts of accuracy—up to 97 percentage points on some methods—on images that have been transmitted over social media or re-digitized, while humans regain much of the lost ac

desk verdict RRDataset is a genuinely useful public stress-test benchmark; the human few-shot conclusion is the one claim I would not defend as written. read the letter →

arxiv 2509.09172 v1 pith:LSVNWZQS submitted 2025-09-11 cs.CV

Chunxiao Li , Xiaoxiao Wang , Meiling Li , Boming Miao , Peng Sun , Yunjian Zhang , Xiangyang Ji , Yao Zhu This is my paper
classification cs.CV
keywords AI-generatedimagedetectionreal-worldrobustnessinternettransmissionre-digitizationhumanfew-shotlearningvision-languagemodelsRRDatasettrustcrisis
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 introduces RRDataset, a benchmark for AI-generated image detection under real-world conditions, covering seven news and everyday scenarios plus images degraded by internet transmission and re-digitization. It claims that all 17 tested detectors and 10 vision-language models lose substantial accuracy on these degraded images, with the largest fake-accuracy drops exceeding 90 percentage points. It further claims that humans also struggle on degraded images but improve notably after a two-image few-shot phase, unlike current detectors. If these claims hold, existing benchmarks overstate detector robustness and human-style adaptation should guide the next generation of detection algorithms.

What carries the argument

The carrying object is RRDataset, a real-world robustness benchmark whose test set is organized along three axes: seven content scenarios (war, disasters, political events, medical/public health, culture/religion, labor, everyday life); multi-round internet transmission across eight platforms; and four re-digitization pipelines (scanning a print, photographing a print, photographing a screen, photographing a projection). The human benchmark uses a two-phase design—120 images before and after viewing two exemplars—to measure few-shot adaptation. These mechanisms together expose performance gaps that artifact-level evaluations miss.

What would settle it

Run the human benchmark with a control group that completes Phase Two without seeing the two exemplar images. If that group improves by roughly the same margin (about 10–15 points on transmitted and re-digitized images), the few-shot-learning conclusion is unsupported.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the robustness of AI-generated image detectors is far lower under realistic image degradations than standard benchmarks suggest. On RRDataset, the best overall accuracy is 89.59%, and internet transmission or re-digitization cuts fake-image accuracy dramatically for most methods—for example, SAFE falls by 97.41% and Freq-Net by 71.61% after transmission, while DIRE and DNF fall by 88.30% and 90.57% after re-digitization. Human observers also lose accuracy on degraded images, but after viewing just two illustrative examples they improve by roughly 10 to 15 percentage points on transmitted and re-digitized images, suggesting a rapid adapt

Load-bearing premise

The claim that humans show few-shot learning assumes the improvement from Phase One to Phase Two is caused by the two exemplar images, not by practice, memory, or growing familiarity with the task—and the study has no control condition that repeats Phase One without showing exemplars.

Editorial extensions

If this is right

  • If the benchmark is right, any detector trained only on clean images will overfit to artifacts that real-world sharing removes; evaluation should include transmitted and re-digitized copies.
  • The near-complete collapse of some frequency- and artifact-based detectors after transmission suggests their features are not stable under compression, while diffusion-based methods resist transmission but fail after re-digitization.
  • AIDE, which combines CLIP-based semantic features, shows the smallest drop under both degradations, indicating that robustness correlates with reliance on content rather than low-level traces.
  • GPT-4o's zero-shot accuracy surpasses most specialized detectors on original images, and a robustness-oriented in-context prompt improves its overall accuracy to 87.47%, approaching the best trained detector.
  • Human few-shot gains imply that a detector given a few labeled examples at deployment time could adapt to a specific degradation, a design direction the paper explicitly urges.

Reading between the lines

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

  • Beyond the paper's reported numbers, the human two-phase improvement may partly reflect practice with the test protocol rather than learning from the two exemplars; a control condition repeating Phase One without exemplars would separate the two effects.
  • The re-digitization protocol could be standardized as a stress test for detector claims, because it is reproducible and it separates methods that use semantic content from those that rely on pixel-level fingerprints.
  • The attribution data—which visual cues participants cited—could be mined to build explainable detectors or to craft prompts for VLMs, turning human judgment criteria into training signal.
  • If the trust-crisis result generalizes, then high accuracy alone may not restore credibility of real images in sensitive domains; deployment may require calibrated uncertainty or provenance labels alongside detection.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces RRDataset, a new benchmark for AI-generated image detection under real-world conditions, comprising 10,000 AI and 10,000 real images across seven content scenarios, with additional versions degraded by multi-round internet transmission and four re-digitization processes. The authors benchmark 17 specialized detectors and 10 vision-language models, reporting substantial drops in fake-image accuracy after transmission and re-digitization (e.g., Freq-Net -71.61, Fusing -79.86, SAFE -97.41 on transmitted images; DIRE -88.30, DNF -90.57 on re-digitized images). They also report a human study with 192 participants and 240 test images, claiming that humans show 'remarkable few-shot learning' after seeing two additional training images, and use this to motivate a robustness-oriented in-context learning prompt for VLMs. The dataset is publicly released and the benchmark is positioned as a more realistic complement to existing datasets such as GenImage and Chameleon.

Significance. If the measurements are reliable, the paper provides a valuable community resource: a public dataset with realistic degradations and a broad detector/VLM comparison. The observed fragility of current detectors under transmission and re-digitization is an important, actionable finding, and the paper is careful to apply transformations symmetrically to real and AI images, so the reported drops are not an artifact of class-asymmetric preprocessing. The human study is large for this literature, and the idea that human judgments improve after brief exposure is interesting. However, the causal claim that the two-image learning phase causes the improvement is not currently identified, and the detector evaluation has an undisclosed train/test split and no variance reporting. These issues are load-bearing for the paper's headline conclusions, especially the human-inspired in-context learning contribution.

major comments (3)
  1. [Sec. 5.1, Sec. 5.2, Table 3] The 'remarkable few-shot learning' claim is confounded. The paper states that Phase Two 'replicated the Phase One procedure' but does not state whether the same 120 test images were reused or new images were presented, and there is no control condition that repeats Phase One without the two learning images. The 10-15 point accuracy increases in Table 3 could therefore be caused by image memorization, task practice, or familiarization with the 14-category reason list and confidence scale, rather than by the two-image learning phase. This matters because Sec. 5.3 explicitly motivates the VLM in-context learning prompt from this result and the conclusion repeats the causal claim. Please provide a control condition or, at minimum, explicitly state that Phase Two used new images and show that the improvement is not explained by simple retest effects.
  2. [Sec. 4.1, Table 2] The detector training/evaluation protocol is underspecified. The text says detectors are 'pretrained on GenImage-SD v1.4 and fine-tuned using RRDataset-subset' and then results are reported on RRDataset, but the paper does not disclose how RRDataset-subset was chosen, whether it is disjoint from the test set, or whether transmitted/re-digitized versions in the test set are derived from the same source images used in fine-tuning. If the subset overlaps with the test source images, the absolute accuracies and cross-detector comparisons in Table 2 are inflated by leakage. This is a core methodological detail for a benchmark paper and must be documented.
  3. [Sec. 4.2, Table 2] Table 2 reports averages of three trials, but no standard deviations, per-trial values, or confidence intervals are given. This is particularly problematic for interpreting small differences that the text describes as 'minor fluctuations' (e.g., DIRE +0.62, DNF +0.36, AIDE -2.91 on transmitted images) versus large drops (e.g., Freq-Net -71.61). Without variance information, the reader cannot judge whether these differences are within run-to-run noise. The same issue applies to the VLM results, which are API-based and non-deterministic. Please report standard deviations and the number of images per cell, and specify what varies across the three trials (e.g., random seeds or API sampling).
minor comments (6)
  1. [Sec. 5.1] The description of the few-shot learning phase is ambiguous: 'Each participant viewed two additional images (original, transmitted, and re-digitized)' lists three conditions but only two images. Please clarify how the two images cover the three conditions, or whether the intended wording is 'two additional images per condition.'
  2. [Fig. 3, Table 4, Sec. 5.3] There are several typos: 'Benckmark' in the Fig. 3 caption, 'Redigital' in the Table 4 header, and 'DCRT-ConvB' in Sec. 5.3 (should be DRCT-ConvB).
  3. [Table 1] Column headings 'Trans-Times Percentage' and 'Self-Trans-Limit' are unclear. The workflow described in the caption should be explained more precisely, including what 'self-trans-limit' means and how the percentages are applied.
  4. [Fig. 1] Figure 1 labels the benchmark 'Real-Robust Bench' while the paper uses 'RRBench' and 'RRDataset.' Please unify the terminology.
  5. [Sec. 5.2] The aggregate numbers in the first paragraph (e.g., 69.17%, 59.52%, 58.29%, 39.64%) are averages over the two phases and the three conditions, but the text does not state this. Table 3 alone does not display these aggregates, which makes the numbers hard to verify.
  6. [Throughout] References to Appendices A-G appear throughout, but the appendices were not included in the version under review. Please ensure the supplementary material is available and cross-referenced correctly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark results are independent measurements; the few-shot human-learning claim has an experimental confound, not a circular reduction.

full rationale

This paper is an empirical benchmark rather than a derivation. The reported detector and VLM accuracies are measurements on images processed by a described transformation pipeline (Sec. 3.3), with labels fixed by original source; no reported number is constructed from the assumptions of the paper. Detectors are pretrained on GenImage and fine-tuned on RRDataset-subset (Sec. 4.1), but this is a standard training/evaluation protocol, not a fitted parameter that is later renamed as a prediction. The human few-shot comparison (Sec. 5.1-5.2) compares Phase One and Phase Two after showing two images; because there is no control arm and the paper does not state whether Phase Two reuses the same 120 images, the causal attribution to two-shot learning is threatened by practice, memory, or familiarization effects. This is an experimental confound and a correctness risk, not circularity: the Phase One-to-Two improvement is not equal by construction to the learning-phase input. Similarly, the in-context learning prompt (Sec. 5.3) was designed after observing human data and evaluated on the same benchmark, which is a mild post hoc selection or overfitting risk, but the reported VLM accuracies are new measurements rather than restatements of the prompt or of the human results. No load-bearing self-citations are present: references to Chameleon, GenImage, FakeBench, and prior detectors are external prior work. Overall, no step in the paper reduces to its own inputs, so the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

This is a benchmark paper rather than a derivation, so the ledger contains the hand-chosen protocol quantities that shape the numbers (transmission mix, re-digitization mix, CLIP filter) and the labeling/experimental assumptions the conclusions rely on.

free parameters (3)
  • CLIP_score_filter_threshold = 0.27
    Used to discard AI images with low text-image alignment during RRDataset construction; changes the difficulty and composition of the AI subset.
  • transmission_cycle_distribution = 2 cycles 10%, 3 cycles 25%, 4 cycles 25%, 5 cycles 25%, 6 cycles 15%
    Hand-defined distribution of transmission rounds across eight platforms; directly sets how much degradation appears in the 'Transmission' test subset and therefore the aggregate drops.
  • re-digitization_method_assignment = equal probability among scanning/printing/screen/projector
    Four re-digitization methods applied with equal probability; this mix determines the character and magnitude of re-digitization accuracy numbers.
assumptions (3)
  • domain assumption The original-source label remains valid after transmission and re-digitization; transformations do not change ground truth.
    Stated in Sec 3.1: 'an image's label (Real vs. AI) is determined solely by its original source.' This is the central labeling assumption; if transformations changed semantic class, the benchmark's ground truth is compromised.
  • domain assumption Fine-tuning detectors on RRDataset-subset and evaluating on the RRDataset test set does not leak identity or transformations between the subsets.
    Sec 4.1 says detectors are pretrained on GenImage-SD v1.4 and fine-tuned on RRDataset-subset; the paper does not disclose how the subset is carved from RRDataset or whether transformed versions of the same images appear in both fine-tuning and test.
  • domain assumption The Phase One to Phase Two improvement in the human study is attributable to the two-shot learning phase rather than practice, memory, or task familiarization.
    Sec 5.1 describes Phase One, a two-image learning phase, then Phase Two, but no control group repeats Phase One without learning images; without this assumption the few-shot learning conclusion is not identified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging the Gap Between Ideal and Real-world Evaluation: Benchmarking AI-Generated Image Detection in Challenging Scenarios." pith.science (2026). https://pith.science/paper/LSVNWZQS

@misc{pith2026250909172,
  author       = {Pith},
  title        = {Pith review of: Bridging the Gap Between Ideal and Real-world Evaluation: Benchmarking AI-Generated Image Detection in Challenging Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LSVNWZQS}},
  note         = {Machine review of arXiv:2509.09172}
}
read the original abstract

With the rapid advancement of generative models, highly realistic image synthesis has posed new challenges to digital security and media credibility. Although AI-generated image detection methods have partially addressed these concerns, a substantial research gap remains in evaluating their performance under complex real-world conditions. This paper introduces the Real-World Robustness Dataset (RRDataset) for comprehensive evaluation of detection models across three dimensions: 1) Scenario Generalization: RRDataset encompasses high-quality images from seven major scenarios (War and Conflict, Disasters and Accidents, Political and Social Events, Medical and Public Health, Culture and Religion, Labor and Production, and everyday life), addressing existing dataset gaps from a content perspective. 2) Internet Transmission Robustness: examining detector performance on images that have undergone multiple rounds of sharing across various social media platforms. 3) Re-digitization Robustness: assessing model effectiveness on images altered through four distinct re-digitization methods. We benchmarked 17 detectors and 10 vision-language models (VLMs) on RRDataset and conducted a large-scale human study involving 192 participants to investigate human few-shot learning capabilities in detecting AI-generated images. The benchmarking results reveal the limitations of current AI detection methods under real-world conditions and underscore the importance of drawing on human adaptability to develop more robust detection algorithms.

Figures

Figures reproduced from arXiv: 2509.09172 by the authors.

Figure 1
Figure 1. Real-Robust Bench overview, featuring a wide range of scenarios and real-world robustness incorporating four re-digitization [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Special-scenario image generation pipeline, illustrating scenario definition, theme expansion, prompt refinement, and final image [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Human Benckmark Evaluation System. Few-Shot Learning Phase. Each participant viewed two additional images (original, transmitted, and re-digitized) drawn from RRDataset, ensuring no overlap with the main test set. Phase Two Testing. Phase Two replicated the Phase One procedure to evaluate whether 2-shot learning affected per￾formance. Comprehensive Inquiry . Finally, two additional questions were posed: 1. When unce… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Trust Crisis Across Everyday Scenarios & Special Sce [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Uncertainty Levels across original images, transmission [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Comparison of Image Attributes Across Original, Transmission and Re-digital. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Frequency-Aware Semantic Fusion with Gated Injection for AI-generated Image Detection

    cs.CV 2026-04 unverdicted novelty 5.0 of 10

    FGINet uses a band-masked frequency encoder and layer-wise gated injection to fuse frequency artifacts with vision foundation model semantics, plus hyperspherical compactness learning, to achieve better generalization...

  2. SPECTRA-Net: Scalable Pipeline for Explainable Cross-domain Tensor Representations for AI-generated Images Detection

    cs.CV 2026-05 unverdicted novelty 4.0 of 10

    SPECTRA-Net fuses multi-view tensor representations from vision foundation models, spectral analysis, local anomaly detection, and statistical descriptors to achieve state-of-the-art cross-domain AI-generated image de...

Reference graph

Works this paper leans on

63 extracted references · 17 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    arXiv

  2. [2]

    Introducing the next generation of claude

    Anthropic. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3- family, 2024. 4, 6

    2024

  3. [3]

    Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966, 1(2):3, 2023

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 4, 6

    arXiv 2023

  4. [4]

    Cifake: Image classifica- tion and explainable identification of ai-generated synthetic images.IEEE Access, 2024

    Jordan J Bird and Ahmad Lotfi. Cifake: Image classifica- tion and explainable identification of ai-generated synthetic images.IEEE Access, 2024. 2

    2024

  5. [5]

    Real-time deepfake detection in the real-world, 2024

    Bar Cavia, Eliahu Horwitz, Tal Reiss, and Yedid Hoshen. Real-time deepfake detection in the real-world, 2024. 2

    2024

  6. [6]

    What makes fake images detectable? understanding proper- ties that generalize

    Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding proper- ties that generalize. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVI 16, pages 103–120. Springer, 2020. 1, 3

    2020

  7. [7]

    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. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558–3568, 2021. 4

    2021

  8. [8]

    DRCT: diffusion reconstruction contrastive training towards universal detection of diffusion generated images

    Baoying Chen, Jishen Zeng, Jianquan Yang, and Rui Yang. DRCT: diffusion reconstruction contrastive training towards universal detection of diffusion generated images. InForty- first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net,

    2024

Show all 63 references
  1. [9]

    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. 1, 4, 6

    2024 arXiv

  2. [10]

    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), pa...

    2023

  3. [11]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 2

    2009

  4. [12]

    Watch your up-convolution: CNN based generative deep neural networks are failing to reproduce spectral distributions

    Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: CNN based generative deep neural networks are failing to reproduce spectral distributions. In 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, CVPR 2020, Seattle, WA, USA, June 13...

    2020

  5. [13]

    Witherden

    Tarik Dzanic, Karan Shah, and Freddie D. Witherden. Fourier spectrum discrepancies in deep network generated images. InAdvances in Neural Information Processing Sys- tems 33: Annual Conference on Neural Information Process- ing Systems 2020, NeurIPS 2020, December 6-12, 2020, ...

    2020

  6. [14]

    Fake-gpt: Detecting fake image via large language model

    Yuming Fan, Dongming Yang, Jiguang Zhang, Bang Yang, and Yuexian Zou. Fake-gpt: Detecting fake image via large language model. InChinese Conference on Pattern Recogni- tion and Computer Vision (PRCV), pages 122–136. Springer,

  7. [15]

    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, 13-18 July 2020, Virtual...

    2020

  8. [16]

    Chatglm: A family of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793, 2024

    Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793, 2024. 4, 6

    2024 arXiv

  9. [17]

    grok-2.https://x.ai/blog/grok-2, 2025

    grok. grok-2.https://x.ai/blog/grok-2, 2025. 4, 6

    2025

  10. [18]

    Wildfake: A large-scale chal- lenging dataset for ai-generated images detection.arXiv preprint arXiv:2402.11843, 2024

    Yan Hong and Jianfu Zhang. Wildfake: A large-scale chal- lenging dataset for ai-generated images detection.arXiv preprint arXiv:2402.11843, 2024. 1, 2

    2024 arXiv

  11. [19]

    hunyuan-vision.https : / / hunyuan

    hunyuan. hunyuan-vision.https : / / hunyuan . tencent.com/, 2025. 4, 6

    2025

  12. [20]

    Fusing global and local features for gen- eralized ai-synthesized image detection

    Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for gen- eralized ai-synthesized image detection. In2022 IEEE In- ternational Conference on Image Processing, ICIP 2022, Bordeaux, France, 16-19 October 2022, pages 3465–3469...

    2022

  13. [21]

    Evolution of detection performance through- out the online lifespan of synthetic images.arXiv preprint arXiv:2408.11541, 2024

    Dimitrios Karageorgiou, Quentin Bammey, Valentin Por- cellini, Bertrand Goupil, Denis Teyssou, and Symeon Pa- padopoulos. Evolution of detection performance through- out the online lifespan of synthetic images.arXiv preprint arXiv:2408.11541, 2024. 2

    2024 arXiv

  14. [22]

    Progressive growing of gans for improved qual- ity, stability, and variation.arXiv preprint arXiv:1710.10196,

    Tero Karras. Progressive growing of gans for improved qual- ity, stability, and variation.arXiv preprint arXiv:1710.10196,

    arXiv

  15. [23]

    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. 4

    2019

  16. [24]

    Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35:26565–26577, 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35:26565–26577, 2022. 1

    2022

  17. [25]

    Harness- ing the power of large vision language models for synthetic image detection.arXiv preprint arXiv:2404.02726, 2024

    Mamadou Keita, Wassim Hamidouche, Hassen Bougueffa, Abdenour Hadid, and Abdelmalik Taleb-Ahmed. Harness- ing the power of large vision language models for synthetic image detection.arXiv preprint arXiv:2404.02726, 2024. 1

    2024 arXiv

  18. [26]

    Clip- ping the deception: Adapting vision-language models for universal deepfake detection

    Sohail Ahmed Khan and Duc-Tien Dang-Nguyen. Clip- ping the deception: Adapting vision-language models for universal deepfake detection. InProceedings of the 2024 International Conference on Multimedia Retrieval, ICMR 2024, Phuket, Thailand, June 10-14, 2024, pages 1006–

    2024

  19. [27]

    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. 3

    2024

  20. [28]

    Flux.https://github.com/ black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/ black-forest-labs/flux, 2024. 3, 4

    2024

  21. [29]

    Improving synthetic image detection to- wards generalization: An image transformation perspective

    Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, and Fuli Feng. Improving synthetic image detection to- wards generalization: An image transformation perspective. arXiv preprint arXiv:2408.06741, 2024. 3, 4, 5, 6

    2024 arXiv

  22. [30]

    Fakebench: Uncover the achilles’ heels of fake images with large multimodal models.arXiv preprint arXiv:2404.13306, 2024

    Yixuan Li, Xuelin Liu, Xiaoyang Wang, Shiqi Wang, and Weisi Lin. Fakebench: Uncover the achilles’ heels of fake images with large multimodal models.arXiv preprint arXiv:2404.13306, 2024. 2, 5

    2024 arXiv

  23. [31]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceeding...

    2014

  24. [32]

    Detecting generated images by real images

    Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIV, pages 95–110. Springer, 2022. 4, 6

    2022

  25. [33]

    Zhengzhe Liu, Xiaojuan Qi, and Philip H. S. Torr. Global texture enhancement for fake face detection in the wild. In 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, CVPR 2020, Seattle, WA, USA, June 13- 19, 2020, pages 8057–8066. Computer Vision Foundati...

    2020

  26. [34]

    Seeing is not always believing: benchmarking human and model perception of ai-generated images.Advances in Neural Information Processing Sys- tems, 36, 2024

    Zeyu Lu, Di Huang, Lei Bai, Jingjing Qu, Chengyue Wu, Xi- hui Liu, and Wanli Ouyang. Seeing is not always believing: benchmarking human and model perception of ai-generated images.Advances in Neural Information Processing Sys- tems, 36, 2024. 1, 2

    2024

  27. [35]

    Detecting gan-generated images by orthogonal training of multiple cnns

    Sara Mandelli, Nicol `o Bonettini, Paolo Bestagini, and Ste- fano Tubaro. Detecting gan-generated images by orthogonal training of multiple cnns. In2022 IEEE International Con- ference on Image Processing, ICIP 2022, Bordeaux, France, 16-19 October 2022, pages 3091–3095. IEEE,...

    2022

  28. [36]

    Midjourney, 2024

    Midjourney. Midjourney, 2024. 4

    2024

  29. [37]

    moonshot-preview-vision.https://www

    moonshoot. moonshot-preview-vision.https://www. moonshot.cn/, 2025. 4, 6

    2025

  30. [38]

    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. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 24480–24489. IEEE, 2023. 1, 3, 4, 6

    2023

  31. [39]

    Semi-truths: A large-scale dataset of ai-augmented images for evaluating robustness of ai-generated image detectors

    Anisha Pal, Julia Kruk, Mansi Phute, Manognya Bhat- taram, Diyi Yang, Duen Horng Chau, and Judy Hoffman. Semi-truths: A large-scale dataset of ai-augmented images for evaluating robustness of ai-generated image detectors. Advances in Neural Information Processing Systems, 37: ...

    2025

  32. [40]

    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,

  33. [41]

    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, p...

    2021

  34. [42]

    Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 4

    2022 arXiv

  35. [43]

    Aer- oblade: Training-free detection of latent diffusion images using autoencoder reconstruction error

    Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aer- oblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130–9140, 2024. 1, 3

    2024

  36. [44]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 3, 4

    2022

  37. [45]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information ...

    2022

  38. [46]

    Forsyth, and Anand Bhattad

    Ayush Sarkar, Hanlin Mai, Amitabh Mahapatra, Svetlana Lazebnik, David A. Forsyth, and Anand Bhattad. Shadows don’t lie and lines can’t bend! generative models don’t know projective geometry...for now.CoRR, abs/2311.17138, 2023. 1, 3

    2023 arXiv

  39. [47]

    De-fake: Detection and attribution of fake images generated by text- to-image generation models

    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 Communica- tions Security, pages 3418–3432, 2023. 1, 2

    2023

  40. [48]

    Learning on gradients: Generalized ar- tifacts representation for gan-generated images detection

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized ar- tifacts representation for gan-generated images detection. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17- ...

    2023

  41. [49]

    C2p-clip: Inject- ing category common prompt in clip to enhance generaliza- tion in deepfake detection.arXiv preprint arXiv:2408.09647,

    Chuangchuang Tan, Renshuai Tao, Huan Liu, Guanghua Gu, Baoyuan Wu, Yao Zhao, and Yunchao Wei. C2p-clip: Inject- ing category common prompt in clip to enhance generaliza- tion in deepfake detection.arXiv preprint arXiv:2408.09647,

    arXiv

  42. [50]

    Frequency-aware deepfake de- tection: Improving generalizability through frequency space domain learning

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake de- tection: Improving generalizability through frequency space domain learning. InProceedings of the AAAI Conference on Artificial Intelligence, pages 5052–5060, 2024. 4, 5, 6

    2024

  43. [51]

    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, ...

    2024

  44. [52]

    Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean- Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023. 4, 6

    2023 arXiv

  45. [53]

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. Cnn-generated images are sur- prisingly easy to spot... for now. In2020 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 8692–

    2020

  46. [54]

    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. InProceedings of the IEEE/CVF international conference on computer vision, pages 22445–22455, 2023. 1, 3, 4, 5, 6

    2023

  47. [55]

    F3net: fusion, feedback and focus for salient object detection

    Jun Wei, Shuhui Wang, and Qingming Huang. F3net: fusion, feedback and focus for salient object detection. InProceed- ings of the AAAI conference on artificial intelligence, pages 12321–12328, 2020. 4, 6

    2020

  48. [56]

    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. 1, 2, 3, 4, 5, 6

    2024 arXiv

  49. [57]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report.arXiv preprint arXiv:2412.15115, 2024. 3

    2024 arXiv

  50. [58]

    Yi: Open foundation models by 01

    Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Guoyin Wang, Heng Li, Jiangcheng Zhu, Jianqun Chen, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 4, 6

    2024 arXiv

  51. [59]

    Detecting and simulating artifacts in GAN fake images

    Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in GAN fake images. InIEEE Inter- national Workshop on Information Forensics and Security, WIFS 2019, Delft, The Netherlands, December 9-12, 2019, pages 1–6. IEEE, 2019. 1

    2019

  52. [60]

    Diffusion noise feature: Ac- curate and fast generated image detection.arXiv preprint arXiv:2312.02625, 2023

    Yichi Zhang and Xiaogang Xu. Diffusion noise feature: Ac- curate and fast generated image detection.arXiv preprint arXiv:2312.02625, 2023. 4, 5, 6

    2023

  53. [61]

    Patchcraft: Exploring texture patch for efficient ai-generated image detection, 2024

    Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, and Xin- peng Zhang. Patchcraft: Exploring texture patch for efficient ai-generated image detection, 2024. 2

    2024

  54. [62]

    Genimage: A million-scale benchmark for de- tecting ai-generated image.Advances in Neural Information Processing Systems, 36, 2024

    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 de- tecting ai-generated image.Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 5

    2024

  55. [8701]

    Computer Vision Foundation / IEEE, 2020. 2, 4, 6

    2020

Pith tools

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