Pith. sign in

REVIEW 3 major objections 5 minor 53 references

A 60,000-image benchmark shows that AI-image detectors near-perfect on known generators drop 4–18 points in accuracy on a single unseen generator.

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

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

T0 review · deepseek-v4-flash

2026-08-02 07:37 UTC pith:YCQ6LZXT

load-bearing objection A useful, well-built resource with one confound that undermines the central OOD architecture claim; the dataset itself deserves engagement but the claims need revision. the 3 major comments →

arxiv 2607.16283 v1 pith:YCQ6LZXT submitted 2026-07-10 cs.CV cs.AI

GenSyn10: A Multi-Generative AI Dataset For Benchmarking Image Classification

Md Faraz Kabir Khan , Saeed Anwar , Ghulam Mubashar Hassan This is my paper
classification cs.CV cs.AI
keywords synthetic image detectionout-of-distribution generalizationbenchmark datasetCIFAR-10deepfake detectiongenerative modelsvision transformersimage classification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces GenSyn10, a 60,000-image benchmark that maps CIFAR-10's class structure onto images from three current, architecturally distinct generative models, while holding out a fourth generator for out-of-distribution testing. Across 17 classifiers, it demonstrates that real-versus-synthetic detection is nearly perfect when the generator is seen (97–99.9% accuracy) but drops by roughly 4–18 points on an unseen generator, even though prompt grammar, resolution, and downsampling are identical. The authors argue that this quantifies the central open problem in AI-image detection—generalization to novel generators—and that transformer-based models with global attention are more resilient than convolutional networks. They frame GenSyn10 as a controlled, reproducible testbed for measuring progress in cross-generator detection and for studying domain adaptation and robustness.

Core claim

The paper's central claim is that the main barrier to detecting AI-generated images is not the detectability of any single generator but the failure of detectors to transfer to an unseen one. On a balanced real-vs-synthetic set built from GenSyn10, all 17 models exceed 97% accuracy in-distribution, most above 99%; against the held-out generator the same models fall to 79–96%, with the largest drops in convolutional and lightweight architectures. The paper attributes the spread to architectural inductive bias rather than capacity, and reports that vision transformers, which use global self-attention, retain the highest OOD accuracy and the smallest drop. It also shows that CIFAR-10-trained cl

What carries the argument

The load-bearing elements are the dataset's construction protocol and its evaluation design. A combinatorial prompt grammar assembles each prompt from six sampled slots plus fixed photorealism constraints, producing over 100,000 unique prompt combinations per class so identical prompts can be run through every generator. A standardized pipeline—30 inference steps at 512×512 resolution, then Lanczos downsampling to 32×32—makes the generator architecture the only variable between in-distribution and out-of-distribution evaluation. The held-out-generator protocol feeds an unused fourth generator the same prompts and post-processing, converting 'detect AI images' into a controlled shift experime

Load-bearing premise

The strongest assumption is that running two of the three main generators with 4-bit quantized weights does not materially change the artifact signatures that classifiers learn, so the benchmark measures generator architecture rather than quantization noise.

What would settle it

Regenerate several hundred images from the two quantized generators at full precision and repeat the binary detection experiment on the same real set; if in-distribution accuracy drops or the OOD gap on the held-out generator narrows materially, quantization artifacts, not architecture, were the learned cues.

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

If this is right

  • Single-generator benchmarks can no longer support claims of general AI-image detection; an OOD rating with a held-out generator is needed to know how a detector will behave in the wild.
  • The 4–18 point accuracy drop, measured with prompts and resolution held fixed, is a concrete baseline number for 17 common architectures on current generators.
  • Transformers and modern hybrids being the most OOD-robust points to global receptive fields and self-attention as a preferred inductive bias for synthetic-image detectors.
  • Because GenSyn10 mirrors CIFAR-10 structure, any classifier or robustness method previously benchmarked on CIFAR-10 can be directly evaluated on the synthetic shift with minimal code changes.
  • The extensible generation pipeline means new generators can be added under the same protocol, keeping the benchmark current as generative models evolve.

Where Pith is reading between the lines

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

  • The persistent high specificity in OOD tests (real-image recall stays near 99%) combined with falling synthetic recall suggests detectors are not losing the ability to rank but are mis-calibrated on OOD scores; a small labeled OOD sample for threshold tuning could recover a large fraction of the accuracy drop—a testable hypothesis the paper does not run.
  • Two of the three main generators were run in 4-bit quantized form; if quantization artifacts dominate what the classifiers learn, the in-distribution results may reflect quantization rather than architecture, and the reported OOD gap could change when the same generators are run at full precision. This can be checked directly by regenerating a subset.
  • The near-perfect semantic classification coexisting with degraded real-vs-synthetic OOD separation implies that the synthetic images are semantically valid yet statistically distinct, suggesting the same dataset could serve as a stress test for generative-model fidelity metrics, not just for detectors.
  • The four-stage protocol (baseline, zero-shot, fine-tune, retention) couples detection quality with catastrophic forgetting; future detectors that use continual-learning or replay strategies could be evaluated on GenSyn10 to see whether OOD robustness can be bought without sacrificing real-data accuracy.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces GenSyn10, a CIFAR-10-aligned synthetic image dataset of 60,000 images generated from three contemporary open-source models (FLUX.2-dev, HunyuanImage-3.0, Qwen-Image-2512), plus a held-out fourth generator (Stable Diffusion 3.5 Large) used only for out-of-distribution (OOD) evaluation. The authors describe a template-based prompt grammar, a standardized generation/resizing pipeline, and a four-stage evaluation protocol (baseline, zero-shot transfer, fine-tuning, retention) applied to 17 classifier architectures. In binary real-vs-synthetic detection, all models perform near-saturation (97–99.9% accuracy) on seen generators but drop to 79–96% on images from the held-out SD 3.5 Large. The paper interprets this drop as evidence of architecture-specific overfitting and argues that transformer/hybrid models are more robust to generator shift.

Significance. If the central claim is valid, GenSyn10 fills a real gap: it provides a modern, multi-generator, controlled benchmark with standardized prompts and a held-out generator protocol, building on earlier single-generator datasets like CIFAKE. The paper's strengths include a reproducible pipeline (deterministic seeds, open-source weights), a broad architectural sweep (17 models), and explicit release of datasets/code/weights. The finding that detectors degrade on an unseen generator is consistent with prior literature and is extended here to recent 2025-generation models. However, the quantitative magnitude of the OOD drop and the architectural-robustness ranking are undermined by a precision confound: two of the three training generators use 4-bit quantization while the held-out generator uses full bfloat16. Because the paper explicitly claims the architecture is the 'sole variable' between ID and OOD evaluation (Appendix B.1), this confound is load-bearing for the main conclusion.

major comments (3)
  1. [§3.2, Table 6, Appendix B.1; central claim in §4.3] The OOD experiment is described as isolating generator architecture: Appendix B.1 states 'the generator architecture is the sole variable differing between in-distribution and OOD evaluation.' Yet Table 6 shows that two of the three training generators (FLUX.2-dev and HunyuanImage-3.0) are run with 4-bit quantization (BnB/NF4), while the held-out SD 3.5 Large is native bfloat16. Thus precision and architecture vary jointly. A real-vs-synthetic classifier can exploit low-level quantization traces—block-wise rounding, altered high-frequency structure—that are common to FLUX and Hunyuan but absent from SD 3.5 Large. The observed 4–18 point drops in Table 4 could therefore reflect a precision shift rather than an architecture-shift generalization failure. The paper provides no ablation (e.g., full-precision FLUX/Hunyuan or a quantized SD 3.5 Large OOD control) to rule out this explanation. A
  2. [§4.3 and Appendix B.1; limitations in §5] The OOD evaluation uses only one held-out generator, SD 3.5 Large, which itself is an MMDiT model—the same paradigm as Qwen-Image-2512, one of the training generators. Therefore the 'unseen' generator is not architecturally novel relative to the training set; differences in text encoders, guidance schemes, and implementation details remain, but these are not the architecture-level shift the paper claims to measure. Moreover, with N=1 OOD generator, rankings such as ViT-Base/16 being the most robust (−3.80 points) and ConvNeXtV2-Base the least (−18.15 points) are single-sample observations. The last paragraph of §5 notes this limitation, but the abstract and contribution 4 present the OOD findings as general. A second held-out generator from a different family (e.g., a GAN or consistency model) is needed to support the broad OOD generalization claim.
  3. [§4.2–4.3, Tables 3–4; contribution 4 and §5] The paper claims that 'architectural inductive bias, not model capacity, is the primary determinant of OOD resilience' (contribution 4, §5). However, the 17 models differ simultaneously in architecture, parameter count, pretraining source, input resolution (32×32 vs 224×224), and fine-tuning hyperparameters. The family-aware protocol (Table 9) gives transformers and hybrids more epochs, lower learning rates, and different warm-up strategies than generic CNNs. Thus the observed cross-family differences in OOD robustness could be due to training schedules or capacity rather than architectural inductive bias. The evidence is correlational and does not include matched-capacity or matched-training-budget comparisons. The claim should be substantially softened or supported by controlled experiments.
minor comments (5)
  1. [General] All experiments use a single random seed (42) with no repeated runs or confidence intervals. Several comparisons in Table 4 involve small gaps (e.g., ViT-Base/16 at −3.80 vs Swin-Base at −4.40); without variance estimates these differences may not be significant. Reporting at least three seeds would strengthen the ranking claims.
  2. [§3.4] When comparing Inception Score to real CIFAR-10 (≈11.2), the authors should specify which real split (train/test) and what number of samples are used, since IS depends on sample size.
  3. [§4.2.1] The text states 'the best model at every stage is transformer-based,' but in Stage 4 the best model is ConvNeXt-Base (99.88%), a hybrid CNN. Rephrase to 'transformer-based or hybrid' for accuracy.
  4. [Table 6 / Appendix B.1] The sentence 'we use quantized weights for generation since our low target resolution makes full-precision inference unnecessary for visual fidelity' (Section 3.2) is an assertion. Even if visual fidelity is preserved, quantization can still alter low-level artifact statistics. Please add a brief justification or cite evidence that 4-bit quantization does not change the detectable trace at 32×32.
  5. [References] Reference [8] (Chatbot Arena) seems to be the wrong citation for the LM Arena Text-to-Image leaderboard; the LM Arena leaderboard is typically cited via lmarena.ai or a separate paper. Please verify.

Circularity Check

0 steps flagged

No circularity: benchmark results are direct empirical measurements; OOD evaluation uses a genuinely held-out generator; no self-citation or fitted-input-as-prediction pattern.

full rationale

GenSyn10 is a dataset/benchmark paper. The four-stage results (Tables 3 and 4) are measured accuracies, F1, and AUC from trained classifiers; no quantity is derived from a fitted parameter and then presented as a prediction. The binary OOD experiment trains on FLUX.2-dev, HunyuanImage-3.0, and Qwen-Image-2512 and evaluates on Stable Diffusion 3.5 Large, which is excluded from GenSyn10 training by construction; the held-out status is stated in Section 3.2 ('we reserve Stable Diffusion 3.5 Large (bfloat16) as a held-out generator, not used to construct GenSyn10'). Thus the 79-96% OOD figures are genuine out-of-sample measurements, not identities. I found no self-citations in the reference list, so no self-citation chain is load-bearing. The claim that 'architectural inductive bias, not model capacity, is the primary determinant of OOD resilience' is a post-hoc interpretation of measured rankings, not a circular derivation. The only flagged issue is a control validity problem, not circularity: Appendix B.1 asserts 'the generator architecture is the sole variable differing between in-distribution and OOD evaluation,' but Table 6 shows FLUX.2-dev and HunyuanImage-3.0 were generated with 4-bit quantization (BnB/NF4) while SD 3.5 Large is native bfloat16. If detectors exploit quantization artifacts, the ID/OOD accuracy gap may partly reflect precision shift rather than architectural shift. This is an untested assumption affecting interpretation of the OOD conclusion; it is a correctness/experimental-design risk, not a reduction of the paper's outputs to its inputs. Separate validation (e.g., generating one training generator at full precision) would resolve it, but its absence does not make the derivation circular.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The paper introduces no fitted free parameters or new postulated entities; it constructs a dataset. The load-bearing assumptions are the representative artifact preservation under downsampling and quantization, and the external validity of the single held-out generator as a proxy for arbitrary unseen generators.

axioms (3)
  • domain assumption Downsampling from 512×512 to 32×32 via Lanczos resampling preserves architecture-specific artifact signatures that detectors can learn.
    Used in Section 3.2 and 3.4; if downsampling erases the subtle artifacts, the detection benchmark loses its signal, but the high zero-shot accuracy suggests semantics survive.
  • domain assumption 4-bit quantized inference (BnB/NF4) on FLUX and Hunyuan produces images with the same characteristic artifacts as their full-precision versions.
    Section 3.2 and Table 6 specify the quantized setups; no ablation against full-precision generation is provided, so the assumption is untested.
  • domain assumption Stable Diffusion 3.5 Large is a representative unseen generator for measuring OOD generalization of detectors.
    Section 4.3 uses SD 3.5 Large as the single held-out generator; the paper acknowledges it shares the MMDiT paradigm with Qwen-Image and cannot characterize all OOD scenarios.

pith-pipeline@v1.3.0-alltime-deepseek · 17242 in / 9256 out tokens · 89480 ms · 2026-08-02T07:37:33.396356+00:00 · methodology

0 comments
read the original abstract

The rapid advancement of generative AI has outpaced our ability to reliably detect its outputs, particularly when detectors encounter generators they have not seen before. We introduce GenSyn10, a CIFAR-10-aligned synthetic image dataset of 60,000 images (10 classes, 32$\times$32, 50k/10k split) generated using three architecturally diverse state-of-the-art models: FLUX.2-dev (Rectified Flow Transformer), HunyuanImage-3.0 (MoE Transformer), and Qwen-Image-2512 (Multimodal Diffusion Transformer), to advance research in AI-generated image detection. A central challenge in this domain is that detectors perform well on known generators but degrade on unseen ones. GenSyn10 addresses this limitation by curating data from multiple contemporary architectures under a standardized generation protocol, enabling controlled and systematic evaluation of out-of-distribution (OOD) generalization to novel generators. Images are generated using a template-based prompt engine and downsampled to ensure consistency. We evaluate 17 image classification models under a four-stage protocol: real-data baseline, zero-shot transfer, fine-tuning, and retention. Despite a measurable domain gap, CIFAR-10-trained models achieve up to 96.86\% zero-shot accuracy on GenSyn10, increasing to 99.88\% after fine-tuning. In binary real-vs-synthetic classification, fine-tuned models achieve 97-99.9\% accuracy on seen generators but drop to 79-96\% on images from an unseen generator, highlighting persistent limitations in OOD generalization. These results establish GenSyn10 as a controlled benchmark for studying synthetic image detection beyond single-generator settings, supporting research on robustness, domain adaptation, and cross-generator generalization.

Figures

Figures reproduced from arXiv: 2607.16283 by Ghulam Mubashar Hassan, Md Faraz Kabir Khan, Saeed Anwar.

Figure 1
Figure 1. Figure 1: Diverse samples from the GenSyn10 dataset illustrating variations within each of its 10 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the benchmark pipeline, including prompt generation, image synthesis [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sample images generated by FLUX.2-dev across the 10 CIFAR-10-aligned categories. [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sample images generated by HunyuanImage-3.0 across the 10 CIFAR-10-aligned categories. [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sample images generated by Qwen-Image-2512 across the 10 CIFAR-10-aligned categories. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sample images generated by Stable Diffusion 3.5 Large across the 10 CIFAR-10-aligned [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Four-stage comparison across core metrics. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

53 extracted references · 16 linked inside Pith

  1. [1]

    Improving image generation with better captions

    James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yilun Guo, et al. Improving image generation with better captions. Technical report, OpenAI, 2023.https://cdn.openai.com/papers/dall-e-3.pdf

    2023

  2. [2]

    CIFAKE: Image classification and explainable identification of AI-generated synthetic images.IEEE Access, 12:15642–15650, 2024

    Jordan J Bird and Ahmad Lotfi. CIFAKE: Image classification and explainable identification of AI-generated synthetic images.IEEE Access, 12:15642–15650, 2024

    2024

  3. [3]

    FLUX.1.https://github.com/black-forest-labs/flux, 2024

    Black Forest Labs. FLUX.1.https://github.com/black-forest-labs/flux, 2024

    2024

  4. [4]

    FLUX.2: Frontier Visual Intelligence

    Black Forest Labs. FLUX.2: Frontier Visual Intelligence. https://bfl.ai/blog/flux-2, 2025

    2025

  5. [5]

    S. Cao, H. Chen, P. Chen, Y . Cheng, Y . Cui, X. Deng, Y . Dong, K. Gong, T. Gu, X. Gu, et al. HunyuanImage 3.0 technical report.arXiv preprint arXiv:2509.23951, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    CO-SPY: Com- bining semantic and pixel features to detect synthetic images by AI

    Siyuan Cheng, Lingjuan Lyu, Zhuo Wang, Xiao Zhang, and Vikash Sehwag. CO-SPY: Com- bining semantic and pixel features to detect synthetic images by AI. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13455–13465, 2025

    2025

  7. [7]

    Pytorch cifar models

    chenyaofo. Pytorch cifar models. https://github.com/chenyaofo/ pytorch-cifar-models, 2023. Accessed: 2025

    2023

  8. [8]

    Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael I

    Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios N. Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael I. Jordan, Joseph E. Gonzalez, and Ion Stoica. Chatbot arena: An open platform for evaluating LLMs by human preference. InInternational Conference on Machine Learning (ICML), 2024

    2024

  9. [9]

    On the detection of synthetic images generated by diffusion models

    Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. InIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2023

    2023

  10. [10]

    Raising the bar of AI-generated image detection with CLIP.arXiv preprint arXiv:2312.00195, 2024

    Davide Cozzolino, Giovanni Poggi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of AI-generated image detection with CLIP.arXiv preprint arXiv:2312.00195, 2024

    Pith/arXiv arXiv 2024

  11. [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. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009

    2009

  12. [12]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  13. [13]

    Scaling rectified flow transformers for high-resolution image synthesis.arXiv preprint arXiv:2403.03206, 2024

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorber, Vikas Shrivastava, and Björn Ommer. Scaling rectified flow transformers for high-resolution image synthesis.arXiv preprint arXiv:2403.03206, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Leveraging frequency analysis for deep fake image recognition

    Joel Frank, Thorsten Eisenhofer, Lea Schönherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. InInternational Confer- ence on Machine Learning (ICML), pages 3247–3258, 2020

    2020

  15. [15]

    Generative adversarial nets.Advances in Neural Information Processing Systems (NeurIPS), 27, 2014

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.Advances in Neural Information Processing Systems (NeurIPS), 27, 2014

    2014

  16. [16]

    Are GAN generated images easy to detect? A critical analysis of the state-of-the-art

    Diego Gragnaniello, Davide Cozzolino, Francesco Marra, Giovanni Poggi, and Luisa Verdoliva. Are GAN generated images easy to detect? A critical analysis of the state-of-the-art. InIEEE International Conference on Multimedia and Expo (ICME), pages 1–6, 2021

    2021

  17. [17]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. 10

    2016

  18. [18]

    GANs trained by a two time-scale update rule converge to a local Nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. InAdvances in Neural Information Processing Systems (NeurIPS), pages 6626–6637, 2017

    2017

  19. [19]

    Denoising diffusion probabilistic models.arXiv preprint arXiv:2006.11239, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.arXiv preprint arXiv:2006.11239, 2020

    Pith/arXiv arXiv 2006

  20. [20]

    Wild- Fake: A large-scale and hierarchical dataset for AI-generated images detection

    Yan Hong, Jiangning Feng, Huan Chen, Jiahui Lan, Hao Zhu, Jie Wang, and Jun Zhang. Wild- Fake: A large-scale and hierarchical dataset for AI-generated images detection. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 39, pages 3500–3508, 2025

    2025

  21. [21]

    Densely connected convolutional networks

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4700–4708, 2017

    2017

  22. [22]

    Progressive growing of GANs for improved quality, stability, and variation

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. InInternational Conference on Learning Representa- tions (ICLR), 2018

    2018

  23. [23]

    A style-based generator architecture for generative adversarial networks.arXiv preprint arXiv:1812.04948, 2019

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks.arXiv preprint arXiv:1812.04948, 2019

    Pith/arXiv arXiv 2019

  24. [24]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009

    2009

  25. [25]

    Improved precision and recall metric for assessing generative models.arXiv preprint arXiv:1904.06991, 2019

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models.arXiv preprint arXiv:1904.06991, 2019

    Pith/arXiv arXiv 1904

  26. [26]

    Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    Pith/arXiv arXiv 2022

  27. [27]

    Swin-base fine-tuned on cifar-10

    Ze Liu. Swin-base fine-tuned on cifar-10. https://huggingface.co/microsoft/ swin-base-patch4-window7-224-in22k , 2024. Fine-tuned on CIFAR-10; Accessed: 2025

    2024

  28. [28]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 10012–10022, 2021

    2021

  29. [29]

    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), pages 11976–11986, 2022

    2022

  30. [30]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

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

    Pith/arXiv arXiv 2017

  31. [31]

    When does label smoothing help? In Advances in Neural Information Processing Systems (NeurIPS), pages 4694–4703, 2019

    Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? In Advances in Neural Information Processing Systems (NeurIPS), pages 4694–4703, 2019

    2019

  32. [32]

    M. Y . Namani, I. Reghioua, G. Bendiab, M. A. Labiod, S. Shiaeles, N. Assem, and A. Al-Nemrat. DeepGuard: Identification and attribution of AI-generated synthetic images.Electronics, 2025

    2025

  33. [33]

    Towards universal fake image detectors that generalize across generative models

    Utkarsh Ojha, Yuheng Li, Jingxiao Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Towards universal fake image detectors that generalize across generative models. arXiv preprint arXiv:2302.10174, 2023

    Pith/arXiv arXiv 2023

  34. [34]

    Pytorch cifar-10 pre-trained models

    Huy Phan. Pytorch cifar-10 pre-trained models. https://github.com/huyvnphan/ PyTorch_CIFAR10, 2021. Accessed: 2025

    2021

  35. [35]

    Qwen-image: Advancing text-to-image generation via latent diffusion

    Qwen Team. Qwen-image: Advancing text-to-image generation via latent diffusion. https: //huggingface.co/Qwen/Qwen-Image-2512, 2025

    2025

  36. [36]

    Design- ing network design spaces.arXiv preprint arXiv:2003.13678, 2020

    Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollár. Design- ing network design spaces.arXiv preprint arXiv:2003.13678, 2020. 11

    Pith/arXiv arXiv 2003

  37. [37]

    Hierarchical text-conditional image generation with CLIP latents.arXiv preprint arXiv:2204.06125, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents.arXiv preprint arXiv:2204.06125, 2022

    Pith/arXiv arXiv 2022

  38. [38]

    AEROBLADE: Training-free detection of latent diffusion images using autoencoder reconstruction error.arXiv preprint arXiv:2401.17879, 2024

    Jonas Ricker, Denis Lukovnikov, and Asja Fischer. AEROBLADE: Training-free detection of latent diffusion images using autoencoder reconstruction error.arXiv preprint arXiv:2401.17879, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022

    2022

  40. [40]

    Improved techniques for training GANs

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. InAdvances in Neural Information Processing Systems (NeurIPS), pages 2234–2242, 2016

    2016

  41. [41]

    MobileNetV2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. MobileNetV2: Inverted residuals and linear bottlenecks. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4510–4520, 2018

    2018

  42. [42]

    DE-FAKE: Detection and attribution of fake images generated by text-to-image generation models.arXiv preprint arXiv:2210.06998, 2023

    Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. DE-FAKE: Detection and attribution of fake images generated by text-to-image generation models.arXiv preprint arXiv:2210.06998, 2023

    Pith/arXiv arXiv 2023

  43. [43]

    Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014

    Pith/arXiv arXiv 2014

  44. [44]

    Stable diffusion 3.5

    Stability AI. Stable diffusion 3.5. https://huggingface.co/stabilityai/ stable-diffusion-3.5-large, 2024

    2024

  45. [45]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1–9, 2015

    2015

  46. [46]

    Re- thinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Re- thinking the inception architecture for computer vision. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818–2826, 2016

    2016

  47. [47]

    MnasNet: Platform-aware neural architecture search for mobile

    Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. MnasNet: Platform-aware neural architecture search for mobile. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2820–2828, 2019

    2019

  48. [48]

    EfficientNet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. InInternational Conference on Machine Learning (ICML), pages 6105–6114, 2019

    2019

  49. [49]

    EfficientNetV2: Smaller models and faster training

    Mingxing Tan and Quoc Le. EfficientNetV2: Smaller models and faster training. InInternational Conference on Machine Learning (ICML), pages 10096–10106, 2021

    2021

  50. [50]

    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. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8695–8704, 2020

    2020

  51. [51]

    ConvNeXt V2: Co-designing and scaling ConvNets with masked autoencoders

    Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. ConvNeXt V2: Co-designing and scaling ConvNets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023

    Pith/arXiv arXiv 2023

  52. [52]

    Vit-base-patch16-224 fine-tuned on cifar-10

    Zhan Wu. Vit-base-patch16-224 fine-tuned on cifar-10. https://huggingface.co/aaraki/ vit-base-patch16-224-in21k-finetuned-cifar10, 2024. Accessed: 2025

    2024

  53. [53]

    GenImage: A million-scale benchmark for detecting AI-generated images.arXiv preprint arXiv:2306.08571, 2024

    Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hao Hu, Jie Hu, and Yunhe Wang. GenImage: A million-scale benchmark for detecting AI-generated images.arXiv preprint arXiv:2306.08571, 2024. 12 Technical appendices and supplementary material A Prompt Grammar Details A.1 Class Instances Table 5 lists the per-class subtyp...

    Pith/arXiv arXiv 2024