Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Few-Shot Learner Generalizes Across AI-Generated Image Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A detector that uses 10 labeled images from an unseen generator outperforms previous state-of-the-art by 11.6% accuracy on the GenImage benchmark.

desk verdict Sensible few-shot framing with a misleading headline number; the zero-shot result is the real contribution. read the letter →

arxiv 2501.08763 v2 pith:GRHFK3LN submitted 2025-01-15 cs.CV

Shiyu Wu , Jing Liu , Jing Li , Yequan Wang This is my paper · ORCID
classification cs.CV
keywords AI-generatedimagedetectionfew-shotlearningprototypicalnetworkdomaingeneralizationdeepfakemetricGendiffusionmodels
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 AI-generated image detection should be reframed as a few-shot classification task: instead of hunting for a universal artifact, a detector should adapt to each new generator using a handful of labeled examples. It introduces Few-Shot Detector (FSD), built on a Prototypical Network, which learns a metric space where images from the same generator cluster together and classifies a query image by its distance to class prototypes. On the GenImage dataset, FSD with 10 shots per unseen generator reaches 84.1% average accuracy, beating the previous best method by 11.6 percentage points. Even without any shots, the zero-shot variant reaches 77.1%, still above the prior state-of-the-art. The practical upshot is that collecting a few samples from a new model at deployment time can substitute for massive retraining.

What carries the argument

The central machinery is the Prototypical Network (Snell et al., 2017) applied to generator-level classes. A ResNet-50 backbone maps each image to a 1024-dimensional embedding. During training, episodes randomly sample 3 classes, 5 support and 5 query images per class, and the loss encourages query embeddings to be close to their class prototype, defined as the mean of support embeddings. At inference, the support set of the unseen generator yields a single prototype for the fake class, real images yield another prototype, and classification is nearest-neighbor with squared Euclidean distance. This removes the need for gradient updates when adapting to a new generator: only a few support embeddings are averaged. The zero-shot variant substitutes prototypes pre-computed from 1024 training images per seen generator, providing a fallback when no samples from the target generator are available.

What would settle it

An experiment that evaluates FSD with 10 support images drawn from a different generator than the query images (e.g., 10 ADM images used to classify BigGAN queries) would test whether the method truly generalizes across generators or instead depends on matching generator identity; if mismatched support reduces accuracy to zero-shot levels, the claim of few-shot generalization across unseen generators is not supported.

Watch

Extended reading notes

Core claim

The central discovery is that a few labeled examples from an unseen generative model suffice to adapt a detector, provided the detector is trained to organize images by generator identity. The authors reconceptualize detection as an N-way K-shot problem where each generator is its own class and real images form a single class. FSD trains a ResNet-50 with episodic Prototypical Network losses, then at test time computes a prototype for the unseen generator as the mean embedding of its support set (e.g., 10 images) and classifies each query by the nearest prototype under squared Euclidean distance. On GenImage, this yields 84.1% average accuracy with 10 shots versus 72.5% for LARE2 and 77.1% for FSD in zero-shot mode. The improvement is roughly saturated at 10 shots: increasing to 200 shots adds only about 2.5% on the ADM class, while decreasing to 1 shot costs about 16.6%. The visualizations indicate that features of unseen generator images form tight clusters, supporting the interpretation that the metric space captures intra-category commonality.

Load-bearing premise

The few-shot advantage rests on the assumption that, at deployment, a user or system can obtain a small labeled sample of real and generated images from the specific new generator being tested.

Editorial extensions

If this is right

  • If the central claim holds, a deployed detector can be updated for a newly released generation model by collecting just 10 labeled images from it, achieving 84.1% average accuracy on GenImage without fine-tuning.
  • The finding that performance saturates near 10 shots implies that the marginal value of extra support samples is small, so the few-shot protocol is cheap to operate in practice.
  • The zero-shot performance of 77.1% indicates that even the episodic multi-class training itself improves generalization over binary training, suggesting that generator-aware structure is broadly beneficial.
  • The success on most test classes suggests that treating generators as separate classes is a viable general strategy, though the VQDM result indicates that very different generators (e.g., those using quantization) may require additional mechanisms.
  • The method's reliance on a metric space may extend to detecting future generators, as long as the learned embeddings preserve intra-class compactness for novel classes.

Reading between the lines

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

  • A plausible operational extension is that platforms could automatically collect a few flagged images from a new generator at release time, enabling near-instant detector updates without full retraining.
  • The behavior on generators with shared backbones (e.g., SD v1.4 and v1.5) suggests that the metric space may not separate models that are architecturally similar, motivating future work on instance-level generator fingerprints.
  • One testable consequence is that the method's few-shot gain should transfer to other media types (e.g., video, audio) if the episodic construction is applied to their feature extractors, which the paper does not explore.
  • The sharp drop on VQDM when excluded from training indicates a possible failure mode: if a generator uses a fundamentally different image representation (such as quantization), the learned metric space may not contain a useful direction for that class, so a confidence-based rejection of the nearest-neighbor decision could be worthwhile.
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 / 5 minor

Summary. The paper proposes Few-Shot Detector (FSD), a prototypical-network method that recasts AI-generated image detection as a few-shot classification task: each generator is treated as a separate class, real images form one class, and a support set of labeled images from an unseen test class is used to compute prototypes for nearest-neighbor classification. The method is trained on a subset of generators and evaluated on held-out generators. On GenImage, Table 1 reports 84.1% average accuracy for FSD with 10 shots versus 72.5% for LARE2, and 77.1% for FSD in the zero-shot setting. The paper also includes ablations on the number of shots and t-SNE visualizations, and it releases code.

Significance. The idea of adapting to unseen generators with a handful of labeled images is practically appealing, since collecting a small labeled sample from a new generator is far cheaper than large-scale dataset construction. The zero-shot result (77.1% average accuracy) also surpasses the strongest baseline LARE2 (72.5%) by 4.6 percentage points, a modest but real improvement. Strengths of the paper include the simple and well-motivated method, the public code release, and the unusual framing of generator identity as a class label. The central weakness is that the headline +11.6% claim compares FSD with access to 10 labeled test-class samples against baselines with no such samples; the method-specific contribution is therefore not identified, and the lack of repeated trials leaves the reported margins potentially within support-set noise.

major comments (3)
  1. [Abstract; Table 1; Section 4.3] The headline claim '+11.6% average accuracy' (Abstract, Table 1) compares FSD with 10 labeled samples from the test class against baselines that receive no test-class samples. Section 4.3 acknowledges that 'our few-shot method may not be directly comparable to other approaches due to accessing only a few samples in the test class,' yet the abstract and conclusion present this as a state-of-the-art result. The reported gain conflates the value of the extra labels with the value of FSD's learned metric space. I request a fair comparison: report FSD zero-shot (77.1%) against all baselines, and add a control baseline that applies the same 10-shot nearest-centroid procedure to a frozen pretrained ResNet-50 feature extractor without the prototypical training, so that the contribution of the learned metric space can be isolated.
  2. [Section 4.2; Table 1] The 10-shot results appear to be based on a single random draw of support images, with no seeds, repeated trials, or variance estimates reported. Because the support set is randomly sampled from each test class (Section 4.2), the per-class accuracies in Table 1 and the curves in Figure 5 could be within support-set sampling noise. Please report the mean and standard deviation over at least 5 independent support-set draws, or at minimum provide fixed seeds and single-run reproducibility details, to establish that the 84.1% vs 77.1% and 84.1% vs 72.5% margins are stable.
  3. [Section 4.6; Figure 5] The 10-shot setting is selected after inspecting the test-class ablation curves: 'we ultimately decide to use the 10-shots results of FSD for comparison' (Section 4.6). Because the same test classes are used to both choose the number of shots and report the headline accuracy, the reported 10-shot numbers carry optimistic selection bias. Please specify a validation-based procedure for choosing the number of shots or report the sensitivity of the main comparison to this hyperparameter choice (e.g., results for 5-shot and 25-shot in Table 1).
minor comments (5)
  1. [Section 4.1] The dataset description says the benchmark contains 8 fake-vs-real subsets, but then the paper merges SD v1.4, SD v1.5, and Wukong into a single 'SD' class, yielding 7 classes; this transition from 8 to 7 should be stated more explicitly to avoid confusion.
  2. [Section 4.3] The protocol for the baseline classifiers in Table 1 is under-specified: 'we report the average performance of 5 classifiers trained on non-test categories and evaluated on each test subset' should state exactly which training sets each baseline used and how the averaging was performed.
  3. [Figure 4] The t-SNE visualizations would be more informative with a legend identifying which color corresponds to the held-out test class in each subfigure, and with the perplexity and number of iterations reported.
  4. [Section 2.2] The related work cites Kukanov et al. on meta-learning for speech deepfakes but does not discuss it in the main text; consider integrating this comparison to position FSD relative to prior few-shot/meta-learning approaches.
  5. [Abstract] The phrase 'without further training' is slightly misleading, since FSD still requires labeled support samples and a forward pass through the network at test time; suggest rewording to 'without fine-tuning' or 'without gradient-based adaptation.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FSD's metric space is trained on held-out generator classes and evaluated on unseen classes; the 10-shot support set is a legitimate test-time input, not a fitted outcome.

full rationale

The paper's derivation chain is self-contained and empirically grounded. FSD trains a Prototypical Network on disjoint generator classes (all but the test subset), and test evaluation uses examples from a class never seen in training. The zero-shot result computes prototypes from training classes only, so it does not use test-class labels. The 10-shot result adds a test-time support set of 10 labeled images from the unseen generator; this is the standard few-shot protocol (Snell et al., 2017), not a fitted parameter renamed as a prediction. The headline +11.6% over LARE2 is an acknowledged asymmetric comparison: Section 4.3 states 'our few-shot method may not be directly comparable to other approaches due to accessing only a few samples in the test class,' so the improvement partly reflects the extra information in the support set. That is a validity and fairness concern, not circularity: no equation reduces to its own input, and no load-bearing claim is justified solely by a self-citation (the paper contains no such self-citation chain). The choice of 10 shots was made after inspecting the ablation curve in Figure 5, which is a hyperparameter selection rather than a derivation that presupposes the reported result.

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

No new physical entities or mediators are introduced. The central claim depends on hyperparameters (number of shots, zero-shot metadata size) and domain assumptions about the availability of few-shot samples and the valid grouping of similar generators. These are standard for an empirical machine learning paper, but the shot count is selected from the ablation and directly affects the headline result.

free parameters (3)
  • number_of_shots = 10
    The number of support samples per unseen class is chosen as the most cost-effective point in the ablation (Figure 5), and the headline +11.6% result uses exactly 10 shots. Selecting the operating point from the ablation is a mild form of fitting to the evaluation setup.
  • training_episode_classes = 3
    Each training episode samples 3 classes from the training set; this is a design hyperparameter, not a physical constant, and the central result depends on it implicitly.
  • zero_shot_metadata_samples = 1024
    The zero-shot metadata vectors are computed from 1024 randomly selected samples per training class; this number is chosen by hand without sensitivity analysis.
assumptions (3)
  • domain assumption Images from unseen generative models can be obtained in real-world scenarios.
    Stated in Section 1: 'We observe that images from unseen domains can actually be obtained in many real-world scenarios.' The entire few-shot setup rests on this premise.
  • domain assumption SD v1.4, SD v1.5, and Wukong share an identical model structure and can be merged into a single class.
    Section 4.1 states these three models share an identical model structure and are merged into one Stable Diffusion subset. This grouping simplifies the task.
  • domain assumption Prototypical Network training on the six training classes transfers to a held-out seventh class.
    The method assumes that the metric space learned from seen generator classes will keep unseen generator classes distinct from real images; this is the standard few-shot transfer assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Few-Shot Learner Generalizes Across AI-Generated Image Detection." pith.science (2026). https://pith.science/paper/GRHFK3LN

@misc{pith2026250108763,
  author       = {Pith},
  title        = {Pith review of: Few-Shot Learner Generalizes Across AI-Generated Image Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GRHFK3LN}},
  note         = {Machine review of arXiv:2501.08763}
}
abstract

Current fake image detectors trained on large synthetic image datasets perform satisfactorily on limited studied generative models. However, these detectors suffer a notable performance decline over unseen models. Besides, collecting adequate training data from online generative models is often expensive or infeasible. To overcome these issues, we propose Few-Shot Detector (FSD), a novel AI-generated image detector which learns a specialized metric space for effectively distinguishing unseen fake images using very few samples. Experiments show that FSD achieves state-of-the-art performance by $+11.6\%$ average accuracy on the GenImage dataset with only $10$ additional samples. More importantly, our method is better capable of capturing the intra-category commonality in unseen images without further training. Our code is available at https://github.com/teheperinko541/Few-Shot-AIGI-Detector.

Figures

Figures reproduced from arXiv: 2501.08763 by the authors.

Figure 1
Figure 1. Challenge facing previous works and our solution. Most classifiers exhibit significant performance degradation over unseen data. We address this challenge by introducing a few-shot strategy which is able to make full use of the limited unseen samples. features from the given samples and use them to generalize across the unseen data, as shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Detection pipeline of Few-Shot Detector. FSD first maps the real and fake images from the support set to the metric space by the Prototypical Network and calculates the representation of each type. A test image is then mapped into the same space and classified according to the nearest-neighbor principle. (Zhang et al., 2019b; Qian et al., 2020; Chai et al., 2020). They treat synthetic image detection as a binary cla… view at source ↗
Figure 3
Figure 3. Prototypical calculation. During training, the metric space is learned by minimizing the distances between query sam￾ples and their corresponding centers. During testing, the test image is only classified between real and fake. ples in each class. xij ∈ R D is an image of D dimensions and yij = i indicates its corresponding label. Each unla￾beled image in the query set will be assigned to one of the N classes. This … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of network features for different classifiers over different classes using t-SNE. Each classifier is trained on those subsets excluding the test one in each sub-figure. The features are extracted from the final layer of our model. distinct categories, whi…
Figure 6
Figure 6. Figure 6: (a) shows that our FSD successfully distinguishes those images across different classes which have been used in training. Images from the unseen ADM class colored (a) FSD (b) Binary Classifier [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

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

    cs.CV 2025-08 conditional novelty 4.0 of 10

    A training-free nearest-neighbor detector built on CLIP intermediate features uses a small number of labeled examples from each new generator to classify deepfakes, reporting strong few-shot accuracy across three benchmarks.

Reference graph

Works this paper leans on

54 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [1]

    In https://www.midjourney.com/home/, 2022

    Midjourney. In https://www.midjourney.com/home/, 2022

    work page 2022

  2. [2]

    In https://xihe.mindspore.cn/modelzoo/wukong, 2022

    Wukong. In https://xihe.mindspore.cn/modelzoo/wukong, 2022

    work page 2022

  3. [3]

    Aunet: Learning relations between action units for face forgery detection

    Bai, W., Liu, Y., Zhang, Z., Li, B., and Hu, W. Aunet: Learning relations between action units for face forgery detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023 , pp.\ 24709--24719. IEEE , 2023

    work page 2023

  4. [4]

    and Cholakov, R

    Boychev, D. and Cholakov, R. Imaginet: A multi-content dataset for generalizable synthetic image detection via contrastive learning. CoRR, abs/2407.20020, 2024

    arXiv 2024

  5. [5]

    Large scale GAN training for high fidelity natural image synthesis

    Brock, A., Donahue, J., and Simonyan, K. Large scale GAN training for high fidelity natural image synthesis. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019

    work page 2019

  6. [6]

    Fakeinversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion

    Cazenavette, G., Sud, A., Leung, T., and Usman, B. Fakeinversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 , pp.\ 10759--10769. IEEE , 2024

    work page 2024

  7. [7]

    What makes fake images detectable? understanding properties that generalize

    Chai, L., Bau, D., Lim, S., and Isola, P. What makes fake images detectable? understanding properties that generalize. In Vedaldi, A., Bischof, H., Brox, T., and Frahm, J. (eds.), Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI , volume 12371 of Lecture Notes in Computer Science, pp.\ 103--12...

    work page 2020

  8. [8]

    A single simple patch is all you need for ai-generated image detection

    Chen, J., Yao, J., and Niu, L. A single simple patch is all you need for ai-generated image detection. CoRR, abs/2402.01123, 2024

    arXiv 2024

Show all 54 references
  1. [9]

    On the detection of synthetic images generated by diffusion models

    Corvi, R., Cozzolino, D., Zingarini, G., Poggi, G., Nagano, K., and Verdoliva, L. On the detection of synthetic images generated by diffusion models. In IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023...

    2023

  2. [10]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L., Li, K., and Fei - Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA , pp.\ 248--255. IEEE Co...

    2009

  3. [11]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Q. Diffusion models beat gans on image synthesis. In Ranzato, M., Beygelzimer, A., Dauphin, Y. N., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems...

    2021

  4. [12]

    Scaling rectified flow transformers for high-resolution image synthesis

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., M \" u ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., Podell, D., Dockhorn, T., English, Z., and Rombach, R. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first Internatio...

    2024

  5. [13]

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

    Goodfellow, I. J., Pouget - Abadie, J., Mirza, M., Xu, B., Warde - Farley, D., Ozair, S., Courville, A. C., and Bengio, Y. Generative adversarial nets. In Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N. D., and Weinberger, K. Q. (eds.), Advances in Neural Information Pro...

    2014

  6. [14]

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

    Gragnaniello, D., Cozzolino, D., Marra, F., Poggi, G., and Verdoliva, L. Are GAN generated images easy to detect? A critical analysis of the state-of-the-art. In 2021 IEEE International Conference on Multimedia and Expo, ICME 2021, Shenzhen, China, July 5-9, 2021 , pp.\ 1--6. ...

    2021

  7. [15]

    Vector quantized diffusion model for text-to-image synthesis

    Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., and Guo, B. Vector quantized diffusion model for text-to-image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 , pp.\ 10686--10696...

    2022

  8. [16]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016 , pp.\ 770--778. IEEE Computer Society, 2016

    2016

  9. [17]

    and Salimans, T

    Ho, J. and Salimans, T. Classifier-free diffusion guidance. CoRR, abs/2207.12598, 2022

    2022 arXiv

  10. [18]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIP...

    2020

  11. [19]

    Analyzing and improving the image quality of stylegan

    Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., and Aila, T. Analyzing and improving the image quality of stylegan. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020 , pp.\ 8107--8116. Computer...

    2020

  12. [20]

    B., Hadid, A., and Taleb - Ahmed, A

    Keita, M., Hamidouche, W., Eutamene, H. B., Hadid, A., and Taleb - Ahmed, A. Bi-lora: A vision-language approach for synthetic image detection. CoRR, abs/2404.01959, 2024

    2024 arXiv

  13. [21]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes. In Bengio, Y. and LeCun, Y. (eds.), 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings , 2014

    2014

  14. [22]

    Meta-learning approaches for improving detection of unseen speech deepfakes

    Kukanov, I., Laakkonen, J., Kinnunen, T., and Hautam \" a ki, V. Meta-learning approaches for improving detection of unseen speech deepfakes. CoRR, abs/2410.20578, 2024

    2024 arXiv

  15. [23]

    Li, J., Li, D., Xiong, C., and Hoi, S. C. H. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In Chaudhuri, K., Jegelka, S., Song, L., Szepesv \' a ri, C., Niu, G., and Sabato, S. (eds.), International Conference on Mach...

    2022

  16. [24]

    Towards robust gan-generated image detection: A multi-view completion representation

    Liu, C., Zhu, T., Shen, S., and Zhou, W. Towards robust gan-generated image detection: A multi-view completion representation. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI 2023, 19th-25th August 2023, Macao, SAR, China , ...

    2023

  17. [25]

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

    Liu, F., Luo, H., Li, Y., Torr, P., and Gu, J. Which model generated this image? A model-agnostic approach for origin attribution. In Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., and Varol, G. (eds.), Computer Vision - ECCV 2024 - 18th European Conference,...

    2024

  18. [26]

    Liu, Z., Qi, X., and Torr, P. H. S. Global texture enhancement for fake face detection in the wild. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020 , pp.\ 8057--8066. Computer Vision Foundation / IEEE , 2020

    2020

  19. [27]

    L., and Keuper, J

    Lorenz, P., Durall, R. L., and Keuper, J. Detecting images generated by deep diffusion models using their local intrinsic dimensionality. In IEEE/CVF International Conference on Computer Vision, ICCV 2023 - Workshops, Paris, France, October 2-6, 2023 , pp.\ 448--459. IEEE , 2023

    2023

  20. [28]

    Lare\( ^ 2 \): Latent reconstruction error based method for diffusion-generated image detection

    Luo, Y., Du, J., Yan, K., and Ding, S. Lare\( ^ 2 \): Latent reconstruction error based method for diffusion-generated image detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 , pp.\ 17006--17015. IEEE , 2024

    2024

  21. [29]

    Exposing the fake: Effective diffusion-generated images detection

    Ma, R., Duan, J., Kong, F., Shi, X., and Xu, K. Exposing the fake: Effective diffusion-generated images detection. CoRR, abs/2307.06272, 2023

    2023 arXiv

  22. [30]

    Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., and Chen, M

    Nichol, A. Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., and Chen, M. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In Chaudhuri, K., Jegelka, S., Song, L., Szepesv \' a ri, C., Niu, G., and Sab...

    2022

  23. [31]

    Ojha, U., Li, Y., and Lee, Y. J. Towards universal fake image detectors that generalize across generative models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023 , pp.\ 24480--24489. IEEE , 2023

    2023

  24. [32]

    T., Rezende, D

    Papamakarios, G., Nalisnick, E. T., Rezende, D. J., Mohamed, S., and Lakshminarayanan, B. Normalizing flows for probabilistic modeling and inference. J. Mach. Learn. Res., 22: 0 57:1--57:64, 2021

    2021

  25. [33]

    L., Li, Y., and Angelov, P

    Pellcier, A. L., Li, Y., and Angelov, P. PUDD: towards robust multi-modal prototype-based deepfake detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 - Workshops, Seattle, WA, USA, June 17-18, 2024 , pp.\ 3809--3817. IEEE , 2024

    2024

  26. [34]

    SDXL: improving latent diffusion models for high-resolution image synthesis

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., M \" u ller, J., Penna, J., and Rombach, R. SDXL: improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Aus...

    2024

  27. [35]

    Thinking in frequency: Face forgery detection by mining frequency-aware clues

    Qian, Y., Yin, G., Sheng, L., Chen, Z., and Shao, J. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In Vedaldi, A., Bischof, H., Brox, T., and Frahm, J. (eds.), Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2...

    2020

  28. [36]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision. In Meila, M. and Zhang, T. (eds.), Proceedings of the 3...

    2021

  29. [37]

    Zero-shot text-to-image generation

    Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., and Sutskever, I. Zero-shot text-to-image generation. In Meila, M. and Zhang, T. (eds.), Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event ...

    2021

  30. [38]

    AEROBLADE: training-free detection of latent diffusion images using autoencoder reconstruction error

    Ricker, J., Lukovnikov, D., and Fischer, A. AEROBLADE: training-free detection of latent diffusion images using autoencoder reconstruction error. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 , pp.\ 9130--9140....

    2024

  31. [39]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 , pp.\ 10674--10685. IEEE , 2022

    2022

  32. [40]

    DE-FAKE: detection and attribution of fake images generated by text-to-image generation models

    Sha, Z., Li, Z., Yu, N., and Zhang, Y. DE-FAKE: detection and attribution of fake images generated by text-to-image generation models. In Meng, W., Jensen, C. D., Cremers, C., and Kirda, E. (eds.), Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Se...

    2023

  33. [41]

    Deepfake-adapter: Dual-level adapter for deepfake detection

    Shao, R., Wu, T., Nie, L., and Liu, Z. Deepfake-adapter: Dual-level adapter for deepfake detection. CoRR, abs/2306.00863, 2023

    2023 arXiv

  34. [42]

    Snell, J., Swersky, K., and Zemel, R. S. Prototypical networks for few-shot learning. In Guyon, I., von Luxburg, U., Bengio, S., Wallach, H. M., Fergus, R., Vishwanathan, S. V. N., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 30: Annual Conference ...

    2017

  35. [43]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net, 2021

    2021

  36. [44]

    C2P-CLIP: injecting category common prompt in CLIP to enhance generalization in deepfake detection

    Tan, C., Tao, R., Liu, H., Gu, G., Wu, B., Zhao, Y., and Wei, Y. C2P-CLIP: injecting category common prompt in CLIP to enhance generalization in deepfake detection. CoRR, abs/2408.09647, 2024 a

    2024 arXiv

  37. [45]

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

    Tan, C., Zhao, Y., Wei, S., Gu, G., Liu, P., and Wei, Y. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. In Wooldridge, M. J., Dy, J. G., and Natarajan, S. (eds.), Thirty-Eighth AAAI Conference on Artificial Intelligence,...

    2024

  38. [46]

    and Hinton, G

    van der Maaten, L. and Hinton, G. Visualizing data using t-sne. Journal of Machine Learning Research, 9 0 (86): 0 2579--2605, 2008

    2008

  39. [47]

    Wang, S., Wang, O., Zhang, R., Owens, A., and Efros, A. A. Cnn-generated images are surprisingly easy to spot... for now. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020 , pp.\ 8692--8701. Computer Vision Fo...

    2020

  40. [48]

    DIRE for diffusion-generated image detection

    Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., and Li, H. DIRE for diffusion-generated image detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023 , pp.\ 22388--22398. IEEE , 2023

    2023

  41. [49]

    Wolleb, J., Bieder, F., Sandk \" u hler, R., and Cattin, P. C. Diffusion models for medical anomaly detection. In Wang, L., Dou, Q., Fletcher, P. T., Speidel, S., and Li, S. (eds.), Medical Image Computing and Computer Assisted Intervention - MICCAI 2022 - 25th International C...

    2022

  42. [50]

    Detecting and simulating artifacts in GAN fake images

    Zhang, X., Karaman, S., and Chang, S. Detecting and simulating artifacts in GAN fake images. In IEEE International Workshop on Information Forensics and Security, WIFS 2019, Delft, The Netherlands, December 9-12, 2019 , pp.\ 1--6. IEEE , 2019 a

    2019

  43. [51]

    Detecting and simulating artifacts in GAN fake images

    Zhang, X., Karaman, S., and Chang, S. Detecting and simulating artifacts in GAN fake images. In IEEE International Workshop on Information Forensics and Security, WIFS 2019, Delft, The Netherlands, December 9-12, 2019 , pp.\ 1--6. IEEE , 2019 b

    2019

  44. [52]

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

    Zhong, N., Xu, Y., Li, S., Qian, Z., and Zhang, X. Patchcraft: Exploring texture patch for efficient ai-generated image detection. CoRR, abs/2311.12397, 2024

    2024 arXiv

  45. [53]

    Genimage: A million-scale benchmark for detecting ai-generated image

    Zhu, M., Chen, H., Yan, Q., Huang, X., Lin, G., Li, W., Tu, Z., Hu, H., Hu, J., and Wang, Y. Genimage: A million-scale benchmark for detecting ai-generated image. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.), Advances in Neural Informatio...

    2023

  46. [54]

    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 10, 2026 · model on record in the stance chip above.