Pith. sign in

REVIEW 4 major objections 6 minor 64 references

Leveraging Failed Samples: A Few-Shot and Training-Free Framework for Generalized Deepfake Detection

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read FTNet claims that a training-free cache of a few labeled images from a new generator, matched by cosine similarity to CLIP intermediate features, beats zero-shot deepfake detectors trained on other generators.

desk verdict A useful empirical baseline in few-shot deepfake detection, but the 'one fake sample' and 'generalized' claims outrun the experiments. read the letter →

arxiv 2508.09475 v1 pith:5SSBJ44O submitted 2025-08-13 cs.CV

Shibo Yao , Renshuai Tao , Xiaolong Zheng , Chao Liang , Chunjie Zhang This is my paper
classification cs.CV
keywords deepfakedetectionfew-shotlearningtraining-freeCLIPfeatureskey-valuecachecosinesimilaritygeneralizableAI-generatedimage
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 argues that deepfake detection should be reframed from a zero-shot generalization problem into a few-shot one, because when a detector fails on a new generator, the very samples it failed on are usually available for use. It proposes FTNet, which builds a key-value cache from a handful of labeled real and fake images of the new generator, embedded with an intermediate CLIP layer, and classifies each test image by cosine similarity to the nearest cached entries. Across 29 generative models on the GenImage, UniversalFakeDetect, and OpenSDI benchmarks, the training-free method reports an average accuracy gain of 8.7% over existing detectors, with a fine-tuned variant FTNet-T gaining 12.1%. The practical consequence is that a detector can be adapted to a newly appearing generator at near-zero cost by exploiting samples that a previous detector already failed on.

What carries the argument

The key-value cache: a Feature Bank of L2-normalized CLIP layer-12 features from the few labeled support images, paired with a Label Bank of one-hot real/fake labels. Classification is a weighted label read-out: cosine similarity $s$ between a test feature and each bank key is mapped through $\exp(-\alpha(1-s))$, and these weights blend the bank labels into the output logits. FTNet-T unfreezes the bank keys, replacing the cosine read-out with a linear layer initialized to the Feature Bank transpose and fine-tuned for 20 epochs.

What would settle it

Repeat the 4-shot protocol but draw the $k$ real and $k$ fake support images 100 times and record the spread of FTNet's accuracy; if the variance is large (say, more than five accuracy points), the claim that a few random samples suffice collapses, since performance would hinge on which samples happen to be drawn. Alternatively, build the support set from a different dataset than the queries (e.g., images of the same generator at a different resolution or compression level) and check whether accuracy falls to the zero-shot baseline.

Watch

Extended reading notes

Core claim

FTNet treats each generative model as its own category and reduces deepfake detection to nearest-label matching in the feature space of a frozen CLIP ViT-L/14 at layer 12. A support set of $k$ real and $k$ fake images per generator is embedded; the L2-normalized features form a Feature Bank and their one-hot real/fake labels form a Label Bank. A test image is embedded the same way, and the cosine similarities to all bank features are turned into weights by $\exp(-\alpha(1-s))$, which blend the cached labels into a real/fake score. The paper reports that this procedure, with no training or parameter updates, achieves state-of-the-art accuracy on three benchmarks spanning GAN, diffusion, and o

Load-bearing premise

The few labeled images randomly sampled from a generator's dataset are representative enough of that generator's full output distribution for nearest-neighbor matching in CLIP space to transfer to the unlabeled test images.

Editorial extensions

If this is right

  • A detector can be deployed for a new generator using only a few labeled examples and no training step, making adaptation far cheaper than retraining or fine-tuning a full detector.
  • The reported accuracy grows with the number of cached samples but with diminishing returns, with 4-shot identified as the practical sweet spot across the three benchmarks.
  • Because matching happens in CLIP feature space, the same cache construction works across GAN, diffusion, and other generator families without architecture changes.
  • Fine-tuning only the cache keys for 20 epochs (FTNet-T) provides a further accuracy boost, so the framework spans both strictly training-free and low-cost-fine-tuning regimes.

Reading between the lines

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

  • A direct consequence the paper does not develop: if CLIP layer-12 features carry generator-specific but semantically invariant artifacts, the same cache trick could be applied to other pre-trained vision encoders, yielding a plug-and-play detector whose backbone can be swapped without re-benchmarking.
  • The benchmarks always draw the support set from the same dataset that later supplies the queries; in deployment the new generator's available samples may come from a different source, resolution, or compression, a distribution shift the reported numbers do not cover.
  • A testable stress test: deliberately draw support samples that are outliers (e.g., the most compressed or lowest-quality images from the generator) and measure the accuracy drop to quantify how dependent the method is on sample representativeness.
  • The 'failed samples' framing suggests an online loop: run a cheap zero-shot detector, collect its false negatives, and feed them into the cache to build a specialized detector for that generator on the fly.
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

4 major / 6 minor

Summary. The paper proposes FTNet, a training-free few-shot framework for AI-generated image detection. It uses a frozen CLIP ViT-L/14 encoder, extracts L2-normalized features from intermediate layer 12, stores k labeled real and fake images per generator in a key-value cache (Feature Bank and Label Bank), and classifies each test image by cosine similarity to the cache, with one-hot labels aggregated under the temperature-scaled activation φ(x)=exp(-α(1-x)); a fine-tuned variant FTNet-T optimizes a linear layer initialized from the cache. Experiments are conducted on GenImage, UniversalFakeDetect, and OpenSDI, with supplementary Chameleon and GANGen results, under 4-shot or 8-shot configurations. The abstract and conclusion claim one fake sample is sufficient and report an average 8.7% improvement over state-of-the-art methods. A more supported statement is that a k-nearest-label retrieval over CLIP L12 features, with k=4–8 labeled examples per generator, accurately classifies in-distribution held-out images from the same generator, and one-shot adaptation works on the Chameleon benchmark.

Significance. If the claims were fully supported, the framework would be practically valuable: it requires no training for the basic variant, adapts to new generators with few labels, and uses a simple, cheap classifier. The paper also contributes broad evaluation across 29 generators and an honest robustness/limitation discussion (Fig. 10, Limitations). However, the contribution as currently stated is narrower: FTNet is a nearest-label cache, and the experimental design gives it labeled target-domain samples that zero-shot baselines do not receive. The useful core — that CLIP intermediate features plus a tiny labeled cache provide a strong few-shot detector — is credible and reproducible in principle, but the headline 'generalized deepfake detection' claim and the 8.7% number require substantial qualification and additional experiments.

major comments (4)
  1. [Abstract; Implementation Details; Tables 1–4] The central claim that FTNet 'uses only one fake sample from an evaluation set' is not supported by the experiments. Implementation Details defines k-shot as sampling k real and k fake images from each generator's dataset, and all main tables are 4-shot or 8-shot; only the supplementary Chameleon experiment uses 1-shot. The abstract's 8.7% improvement and the conclusion therefore refer to a 4/8-shot regime, not one sample. Please either report 1-shot accuracy on all benchmarks or revise the claim. This matters because the one-sample capability is the paper's motivating premise.
  2. [Implementation Details; Tables 1–4; Fig. 10] The method is evaluated only in-distribution. The support cache is built by random sampling from the same generative model datasets that provide the query images, so the reported accuracies show few-shot adaptation to the benchmark distribution, not generalization to 'failed samples' from a different source. The paper's own robustness evaluation (Fig. 10) shows that accuracy on Chameleon degrades under JPEG compression and Gaussian blur, exactly the perturbations that would arise in real deployment. To support the real-world generalization claim, the authors should add a protocol where the support set and query set differ in generator version, resolution, compression, or post-processing (or at least include a clear train/test split by generation configuration).
  3. [Abstract; Tables 1–4; Table 5] The headline 8.7% average improvement is not defined or computable from the tables, and the comparison is not apples-to-apples. In Tables 1–4 the baseline detectors are zero-shot (trained on SDv1.4/ProGAN/SD1.5 with no target labels) while FTNet receives labeled target-domain support; under this protocol a gain is structurally expected. The reported margins are much smaller than 8.7% relative to the strongest baselines: e.g., +1.9% on GenImage cross-domain, +1.03% on UFD, and −2.04% on OpenSDI against MaskCLIP. The proper comparisons are the few-shot fine-tuned baselines in Table 5, where FTNet does well, and the zero-shot comparison in Table 6. Please report the exact aggregation for the 8.7% figure and clearly label the comparison protocol in the abstract.
  4. [Implementation Details; Tables 1–4] No measure of variance is provided for the random support-set sampling. With k=4 or 8 images per generator, accuracy is likely sensitive to which images are cached; the paper reports point estimates with no seed, repeats, or confidence intervals. Please report mean±std over multiple random support draws (or at least a few seeds) for the main tables.
minor comments (6)
  1. [Eqs. (3)–(4); Fig. 3] Notation is inconsistent: F_B and L_B appear as 'FB' and 'LB' in Fig. 3, and the shape of L_B is not stated. Define all dimensions explicitly.
  2. [Table 4] The column header 'A VG' appears to be a typo. Also, Table 5 uses 'mAcc' for all datasets, but the text only defines Accuracy as the primary metric.
  3. [Table 6] The 'Training samples' column reports 500 for FTNet/FTNet-T, but the method description says only one fake and one real sample. Clarify what the 500 images are (likely 250 real + 250 fake) and how this relates to the k-shot definition.
  4. [Table 2 caption] The table mixes zero-shot and few-shot baselines (FSD zero-shot vs FSD 10-shot) without distinguishing them in the caption. Add a footnote or column indicating which baselines have access to target-domain labels.
  5. [Figure 4] The right panel (CLIP layer comparison) has no labeled y-axis; exact accuracy values cannot be read from the figure. Add axis labels or a table.
  6. [Introduction] The introduction mentions FTNet-T achieving 12.1% improvement, but the abstract only reports 8.7%. Explain this discrepancy and define the aggregation for both numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; FTNet is a transparent nearest-label cache; evaluation-design concerns are external-validity issues, not circularity.

full rationale

FTNet's decision rule is stated as Eqs. (3)-(4): features are L2-normalized CLIP intermediate features; logits are φ(S_cos L_B), i.e., a temperature-softened weighted average of one-hot labels keyed by cosine similarity. Nothing in this chain defines the predicted class in terms of itself or fits a parameter to the quantity being predicted. The cache labels are inputs; the output is a similarity-weighted vote over them. There is no self-citation chain or imported uniqueness theorem: Tip-Adapter (Zhang et al. 2021) is cited for the activation and prior same-author works (e.g., NPR, C2P-CLIP) are used only as comparison baselines. The main legitimate concerns are about experimental framing, not circularity: the support set is sampled from the same generative-model datasets as the query set ('we randomly sample k real and k fake images from each generative model’s dataset'), so the reported 8.7% gain over zero-shot baselines largely reflects access to target-domain labels and in-distribution support; and the paper's own robustness analysis concedes that the similarity cache 'is greatly affected by compression or blurring.' Also, the abstract's 'one fake sample' claim is not the configuration used in the main tables (4-/8-shot). These are validity/consistency questions and should be weighed as correctness risk, not as circular derivation. Accordingly, no circular step can be quoted with a specific equation-to-equation or parameter-to-prediction reduction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method is an empirical nearest-neighbor classifier; its free parameters are a temperature, a chosen CLIP layer, and the support-set size, all selected using evaluation performance. Its main axioms are domain assumptions about CLIP feature clustering and the representativeness of target-domain samples. No invented entities are introduced.

free parameters (3)
  • alpha (temperature in activation function) = 15
    Tuning hyperparameter in Eq. 4 that controls sharpness of similarity-to-label weighting; set to 15 without an ablation or validation procedure shown.
  • CLIP layer index = 12 (of 24)
    Selected based on layer ablation (Figure 4 right) because it gave the highest average accuracy on the evaluation datasets; this is a posteriori selection on the test benchmark.
  • Support shot count k = 4 (and 8 in some experiments)
    Number of real and fake images sampled per generator to build the cache; 4-shot chosen from ablation as best tradeoff, a decision made on evaluation data.
assumptions (4)
  • domain assumption CLIP intermediate features encode stable, generator-specific forgery artifacts that cluster separately from real images
    The entire method rests on this; Section 'Generalization Ability of CLIP Feature Layers' and Figure 5 claim it, but it is assumed, not derived.
  • domain assumption Support images sampled from the same datasets as query images are representative of the full target generator distribution
    Implied by cache construction in Section 'Cache Model Construction'; if a new sample is unrepresentative (different source or post-processing), the nearest-neighbor cache has no guarantee.
  • domain assumption Real images from all sources can be treated as a single class and the cache class balance should match real-world mixing
    They unify real images into one category and balance cache counts; this is a modeling choice in the problem definition.
  • domain assumption Treating each generative model as its own class and then merging to real/fake in classification is a valid proxy for binary deepfake detection
    Problem Definition treats generators as categories to build the cache, then uses one-hot real/fake labels; this assumes per-generator prototype structure helps binary separation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Failed Samples: A Few-Shot and Training-Free Framework for Generalized Deepfake Detection." pith.science (2026). https://pith.science/paper/5SSBJ44O

@misc{pith2026250809475,
  author       = {Pith},
  title        = {Pith review of: Leveraging Failed Samples: A Few-Shot and Training-Free Framework for Generalized Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5SSBJ44O}},
  note         = {Machine review of arXiv:2508.09475}
}
read the original abstract

Recent deepfake detection studies often treat unseen sample detection as a ``zero-shot" task, training on images generated by known models but generalizing to unknown ones. A key real-world challenge arises when a model performs poorly on unknown samples, yet these samples remain available for analysis. This highlights that it should be approached as a ``few-shot" task, where effectively utilizing a small number of samples can lead to significant improvement. Unlike typical few-shot tasks focused on semantic understanding, deepfake detection prioritizes image realism, which closely mirrors real-world distributions. In this work, we propose the Few-shot Training-free Network (FTNet) for real-world few-shot deepfake detection. Simple yet effective, FTNet differs from traditional methods that rely on large-scale known data for training. Instead, FTNet uses only one fake samplefrom an evaluation set, mimicking the scenario where new samples emerge in the real world and can be gathered for use, without any training or parameter updates. During evaluation, each test sample is compared to the known fake and real samples, and it is classified based on the category of the nearest sample. We conduct a comprehensive analysis of AI-generated images from 29 different generative models and achieve a new SoTA performance, with an average improvement of 8.7\% compared to existing methods. This work introduces a fresh perspective on real-world deepfake detection: when the model struggles to generalize on a few-shot sample, leveraging the failed samples leads to better performance.

Figures

Figures reproduced from arXiv: 2508.09475 by the authors.

Figure 1
Figure 1. Comparison of traditional methods and our few [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of traditional detector and our method [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overall framework of our proposed FTNet. The core of FTNet extracts features from the intermediate layers of the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Analysis of the method’s performance. (Left) In [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: The top and bottom rows display two query cases [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 5
Figure 5. Figure 5: Visualization of FTNet Using CLIP ViT-L/14 Fea [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Performance comparison with various advanced [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: The visualization of CAM, the top and bottom [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Logit distributions of extracted forgery features. We tested on datasets such as ADM, BigGAN, glide, and SD. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Robustness on JPEG compression and Gaussian [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 40 canonical work pages

  1. [1]

    Brock, A.; Donahue, J.; and Simonyan, K. 2018. Large scale GAN training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096

    arXiv 2018

  2. [2]

    Chai, L.; Bau, D.; Lim, S.-N.; and Isola, P. 2020. What makes fake images detectable? understanding properties that generalize. In European conference on computer vision, 103--120. Springer

    work page 2020

  3. [3]

    Chen, C.; Chen, Q.; Xu, J.; and Koltun, V. 2018. Learning to see in the dark. In Proceedings of the IEEE conference on computer vision and pattern recognition, 3291--3300

    work page 2018

  4. [4]

    Chen, Q.; and Koltun, V. 2017. Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, 1511--1520

    work page 2017

  5. [5]

    Chen, X.; Dong, C.; Ji, J.; Cao, J.; and Li, X. 2021. Image manipulation detection by multi-view multi-scale supervision. In Proceedings of the IEEE/CVF international conference on computer vision, 14185--14193

    work page 2021

  6. [6]

    Choi, Y.; Choi, M.; Kim, M.; Ha, J.-W.; Kim, S.; and Choo, J. 2018. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 8789--8797

    work page 2018

  7. [7]

    Dai, T.; Cai, J.; Zhang, Y.; Xia, S.-T.; and Zhang, L. 2019. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11065--11074

    2019

  8. [8]

    Dhariwal, P.; and Nichol, A. 2021. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34: 8780--8794

    2021

Show all 64 references
  1. [9]

    Durall, R.; Keuper, M.; and Keuper, J. 2020. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 7890--7899

    2020

  2. [10]

    Esser, P.; Kulal, S.; Blattmann, A.; Entezari, R.; M \"u ller, J.; Saini, H.; Levi, Y.; Lorenz, D.; Sauer, A.; Boesel, F.; et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning

    2024

  3. [11]

    Frank, J.; Eisenhofer, T.; Sch \"o nherr, L.; Fischer, A.; Kolossa, D.; and Holz, T. 2020. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, 3247--3258. PMLR

    2020

  4. [12]

    J.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y

    Goodfellow, I. J.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y. 2014. Generative adversarial nets. Advances in neural information processing systems, 27

    2014

  5. [13]

    Gu, S.; Chen, D.; Bao, J.; Wen, F.; Zhang, B.; Chen, D.; Yuan, L.; and Guo, B. 2022. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10696--10706

    2022

  6. [14]

    Guillaro, F.; Zingarini, G.; Usman, B.; Sud, A.; Cozzolino, D.; and Verdoliva, L. 2025. A bias-free training paradigm for more general ai-generated image detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, 18685--18694

    2025

  7. [15]

    He, K.; Chen, X.; Xie, S.; Li, Y.; Doll \'a r, P.; and Girshick, R. 2022. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 16000--16009

    2022

  8. [16]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 6840--6851

    2020

  9. [17]

    Karras, T.; Aila, T.; Laine, S.; and Lehtinen, J. 2017. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196

    2017 arXiv

  10. [18]

    Karras, T.; Laine, S.; and Aila, T. 2019. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4401--4410

    2019

  11. [19]

    P.; and Ba, J

    Kingma, D. P.; and Ba, J. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  12. [20]

    Korshunov, P.; Jain, A.; and Marcel, S. 2022. Custom attribution loss for improving generalization and interpretability of deepfake detection. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 8972--8976. IEEE

    2022

  13. [21]

    Koutlis, C.; and Papadopoulos, S. 2024. Leveraging representations from intermediate encoder-blocks for synthetic image detection. In European Conference on Computer Vision, 394--411. Springer

    2024

  14. [22]

    Labs, B. F. 2024. FLUX. https://github.com/black-forest-labs/flux

    2024

  15. [23]

    Li, K.; Zhang, T.; and Malik, J. 2019. Diverse image synthesis from semantic layouts via conditional imle. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 4220--4229

    2019

  16. [24]

    Liu, H.; Tan, Z.; Tan, C.; Wei, Y.; Wang, J.; and Zhao, Y. 2024. Forgery-aware adaptive transformer for generalizable synthetic image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10770--10780

    2024

  17. [25]

    Liu, Z.; Qi, X.; and Torr, P. H. 2020. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 8060--8069

    2020

  18. [26]

    Luo, Y.; Du, J.; Yan, K.; and Ding, S. 2024. LaRE\^ 2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 17006--17015

    2024

  19. [27]

    Ma, X.; Du, B.; Jiang, Z.; Hammadi, A. Y. A.; and Zhou, J. 2023. IML-ViT: Benchmarking image manipulation localization by vision transformer. arXiv preprint arXiv:2307.14863

    2023 arXiv

  20. [28]

    Marra, F.; Gragnaniello, D.; Verdoliva, L.; and Poggi, G. 2019. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), 506--511. IEEE

    2019

  21. [29]

    Midjourney . 2022. Available at https://www.midjourney.com/home/

    2022

  22. [30]

    Nichol, A.; Dhariwal, P.; Ramesh, A.; Shyam, P.; Mishkin, P.; McGrew, B.; Sutskever, I.; and Chen, M. 2021 a . Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741

    2021 arXiv

  23. [31]

    Q.; and Dhariwal, P

    Nichol, A. Q.; and Dhariwal, P. 2021 b . Improved denoising diffusion probabilistic models. In International conference on machine learning, 8162--8171. PMLR

    2021

  24. [32]

    Ojha, U.; Li, Y.; and Lee, Y. J. 2023. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 24480--24489

    2023

  25. [33]

    Park, T.; Liu, M.-Y.; Wang, T.-C.; and Zhu, J.-Y. 2019. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2337--2346

    2019

  26. [34]

    Podell, D.; English, Z.; Lacey, K.; Blattmann, A.; Dockhorn, T.; M \"u ller, J.; Penna, J.; and Rombach, R. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952

    2023 arXiv

  27. [35]

    Qian, Y.; Yin, G.; Sheng, L.; Chen, Z.; and Shao, J. 2020. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, 86--103. Springer

    2020

  28. [36]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PmLR

    2021

  29. [37]

    Ramesh, A.; Pavlov, M.; Goh, G.; Gray, S.; Voss, C.; Radford, A.; Chen, M.; and Sutskever, I. 2021. Zero-shot text-to-image generation. In International conference on machine learning, 8821--8831. Pmlr

    2021

  30. [38]

    Ricker, J.; Lukovnikov, D.; and Fischer, A. 2024. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9130--9140

    2024

  31. [39]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022 a . High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684--10695

    2022

  32. [40]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022 b . High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 10684--10695

    2022

  33. [41]

    Rossler, A.; Cozzolino, D.; Verdoliva, L.; Riess, C.; Thies, J.; and Nie ner, M. 2019. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, 1--11

    2019

  34. [42]

    Shiohara, K.; and Yamasaki, T. 2022. Detecting Deepfakes with Self-Blended Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 18720--18729

    2022

  35. [43]

    Snell, J.; Swersky, K.; and Zemel, R. 2017. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30

    2017

  36. [44]

    Tan, C.; Tao, R.; Liu, H.; Gu, G.; Wu, B.; Zhao, Y.; and Wei, Y. 2025. C2p-clip: Injecting category common prompt in clip to enhance generalization in deepfake detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 7184--7192

    2025

  37. [45]

    Tan, C.; Tao, R.; Liu, H.; and Zhao, Y. 2024 a . GANGen-Detection: A Dataset generated by GANs for Generalizable deepfake Detection. https://github.com/chuangchuangtan/GANGen-Detection

    2024

  38. [46]

    Tan, C.; Zhao, Y.; Wei, S.; Gu, G.; Liu, P.; and Wei, Y. 2024 b . Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 5052--5060

    2024

  39. [47]

    Tan, C.; Zhao, Y.; Wei, S.; Gu, G.; Liu, P.; and Wei, Y. 2024 c . Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 28130--28139

    2024

  40. [48]

    Tan, C.; Zhao, Y.; Wei, S.; Gu, G.; and Wei, Y. 2023. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12105--12114

    2023

  41. [49]

    Tao, R.; Tan, C.; Liu, H.; Wang, J.; Qin, H.; Chang, Y.; Wang, W.; Ni, R.; and Zhao, Y. 2025. SAGNet: Decoupling Semantic-Agnostic Artifacts from Limited Training Data for Robust Generalization in Deepfake Detection. IEEE Transactions on Information Forensics and Security

    2025

  42. [50]

    Wang, J.; Wu, Z.; Chen, J.; Han, X.; Shrivastava, A.; Lim, S.-N.; and Jiang, Y.-G. 2022. Objectformer for image manipulation detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2364--2373

    2022

  43. [51]

    Wang, S.-Y.; Wang, O.; Zhang, R.; Owens, A.; and Efros, A. A. 2020. CNN-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 8695--8704

    2020

  44. [52]

    Wang, Y.; Huang, Z.; and Hong, X. 2025. OpenSDI: Spotting Diffusion-Generated Images in the Open World. In Proceedings of the Computer Vision and Pattern Recognition Conference, 4291--4301

    2025

  45. [53]

    Wang, Z.; Bao, J.; Zhou, W.; Wang, W.; Hu, H.; Chen, H.; and Li, H. 2023. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 22445--22455

    2023

  46. [54]

    Wu, S.; Liu, J.; Li, J.; and Wang, Y. 2025. Few-Shot Learner Generalizes Across AI-Generated Image Detection. arXiv preprint arXiv:2501.08763

    2025 arXiv

  47. [55]

    Wukong. 2022. Available at https://xihe.mindspore.cn/modelzoo/wukong

    2022

  48. [56]

    Yan, S.; Li, O.; Cai, J.; Hao, Y.; Jiang, X.; Hu, Y.; and Xie, W. 2024 a . A sanity check for ai-generated image detection. arXiv preprint arXiv:2406.19435

    2024 arXiv

  49. [57]

    Yan, Z.; Wang, J.; Jin, P.; Zhang, K.-Y.; Liu, C.; Chen, S.; Yao, T.; Ding, S.; Wu, B.; and Yuan, L. 2024 b . Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection. arXiv preprint arXiv:2411.15633

    2024 arXiv

  50. [58]

    Yang, Y.; Qian, Z.; Zhu, Y.; Russakovsky, O.; and Wu, Y. 2025. D\^ 3: Scaling Up Deepfake Detection by Learning from Discrepancy. In Proceedings of the Computer Vision and Pattern Recognition Conference, 23850--23859

    2025

  51. [59]

    Zhang, R.; Fang, R.; Zhang, W.; Gao, P.; Li, K.; Dai, J.; Qiao, Y.; and Li, H. 2021. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930

    2021 arXiv

  52. [60]

    Zhang, X.; Karaman, S.; and Chang, S.-F. 2019. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), 1--6. IEEE

    2019

  53. [61]

    Zhou, B.; Khosla, A.; Lapedriza, A.; Oliva, A.; and Torralba, A. 2016. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2921--2929

    2016

  54. [62]

    Zhu, J.-Y.; Park, T.; Isola, P.; and Efros, A. A. 2017. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2223--2232

    2017

  55. [63]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  56. [64]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.