Pith. sign in

REVIEW 4 major objections 6 minor 28 references

Swin Transformer for Robust CGI Images Detection: Intra- and Inter-Dataset Analysis across Multiple Color Spaces

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims a Swin Transformer trained on raw RGB pixels separates computer-generated imagery from authentic photos with 95–98% accuracy across three datasets, and that this makes it a strong generic CGI detector.

desk verdict Claims inter-dataset generalization but never actually tests it; the 97-99% headline also clashes with the paper's own Columbia numbers. read the letter →

arxiv 2505.16253 v1 pith:L57YJFJK submitted 2025-05-22 cs.CV

Preeti Mehta , Aman Sagar , Suchi Kumari This is my paper · ORCID
classification cs.CV
keywords SwinTransformercomputer-generatedimagedetectiondigitalforensicscolorspaceanalysisdomaingeneralizationdeeplearningclassificationCGI
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

This paper sets out to show that a Swin Transformer—a hierarchical vision transformer that processes images in shifted local windows—can reliably tell computer-generated imagery (CGI) from authentic photographs. The authors test the model on three datasets: CiFAKE, JSSSTU, and Columbia, across RGB, YCbCr, and HSV color encodings, and report that RGB gives the best accuracy: 98% on CiFAKE, 97% on JSSSTU, 95% on Columbia, and 98% on a pooled combination of all three. They position this as evidence of domain generalization and as a state-of-the-art result, since the Swin model matches ResNet-50 and beats VGG-19 on the pooled RGB test. If correct, the result matters because a single vision transformer could serve as a practical forensic tool for flagging synthetic images without handcrafted features.

What carries the argument

The load-bearing object is the Swin Transformer, a vision transformer whose self-attention operates on non-overlapping local windows that shift between consecutive layers, producing a hierarchical feature map that captures both fine texture and long-range structure. The authors feed it 224×224 images in each of three color spaces (RGB, YCbCr, HSV) and train a binary classifier with cross-entropy loss on the resulting features. Two supporting mechanisms carry the generalization argument: data augmentation to rebalance the severely depleted Columbia subset (43 CGI, 150 real images), and pooling of all three datasets into a balanced 1500-images-per-class set on which the headline accuracy is measured.

What would settle it

Train the same Swin configuration on CiFAKE and JSSSTU only, then test on the un-augmented Columbia CGI and real images; if accuracy falls toward chance or well below the reported 95–98%, the inter-dataset generalization claim fails.

Watch

Extended reading notes

Core claim

The central claim, stated sympathetically, is that the inductive biases of the Swin Transformer—self-attention computed within shifted local windows, giving it both local texture sensitivity and global context—make it particularly suited to the CGI-versus-natural task. The paper reports that RGB frames alone separate the two classes in t-SNE projections more cleanly than YCbCr or HSV, and that the network achieves 0.98 accuracy/precision/recall/F1/AUC on the pooled dataset, matching ResNet-50 and exceeding VGG-19. The discovery is an extension result: an off-the-shelf vision transformer, trained on simple color frames, transfers across three different CGI sources at high accuracy, suggesting that synthetic-image detection may not need bespoke noise or frequency features.

Load-bearing premise

The generalization claim rests on the assumption that a random train/test split of the pooled three-dataset collection predicts performance on an unseen image source, rather than just measuring the model's ability to fit a mixture of known sources.

Editorial extensions

If this is right

  • If the claim holds, a forensic examiner could deploy a single Swin Transformer, pretrained on ordinary classification tasks and fine-tuned on RGB frames, as a first-pass CGI detector across heterogeneous image sources.
  • The reported RGB superiority implies that color-space engineering is unnecessary for this task; the raw pixel distribution in RGB already contains the discriminative signal, at least for these three sources.
  • The pooled-dataset result suggests that mixing CGI sources during training improves generalization, which would argue for building larger, multi-renderer training corpora rather than tuning per-source detectors.
  • Because the Swin model matches ResNet-50, the result indicates that attention-based architectures are at least as effective as CNNs for CGI detection, opening a straightforward architectural path for future forensics models.
  • The data-augmentation recovery of the tiny Columbia subset suggests that even severely imbalanced forensic collections can yield usable classifiers, though the confidence intervals on those numbers are not reported.

Reading between the lines

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

  • The paper calls its pooled train/test split 'inter-dataset testing,' but a random split of the combined pool does not measure transfer to an unseen source; a true domain-shift test would train on, say, CiFAKE and JSSSTU and test on Columbia. That experiment is the real test of the generalization claim.
  • The Columbia results rest on only 43 positive and 150 negative images after data loss, so the 95% accuracy number should be read as testimony about a small augmented sample, not about the full 800-per-class Columbia dataset.
  • The claim that RGB is universally the most discriminative color space is likely source-dependent; a testable extension would run the same architecture on rendered images with unusual chromatic distributions (e.g., stylized game engines or path-traced animation) to see whether HSV or YCbCr becomes competitive.
  • If the Swin features are truly generic, the same model should also separate newer generative outputs (GANs, diffusion models) from real photos; the paper does not test that, but the CIFAKE dataset, which it uses, does contain AI-generated images, so the boundary between CGI and other synthetic imagery is already fuzzy.
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 a Swin Transformer-based binary classifier to distinguish computer-generated imagery (CGI) from natural images, evaluated on three datasets (CiFAKE, JSSSTU, Columbia) across RGB, YCbCr, and HSV color spaces. The authors report per-dataset results (Tables 4–6), a combined-dataset experiment (Table 7), t-SNE visualizations, and a comparison against VGG-19 and ResNet-50 on the combined RGB data (Table 8). The central claim is that the proposed method achieves state-of-the-art accuracy, 'consistently between 97-99% across multiple datasets,' and that inter-dataset testing demonstrates robustness and domain generalization.

Significance. If the central claims were supported, the work would provide a useful benchmark for CGI detection with a modern transformer backbone and a systematic comparison of color spaces. The paper's strengths include the use of three publicly available datasets, consideration of multiple color spaces, and a straightforward comparison with two CNN baselines. However, the claimed inter-dataset generalization is not actually measured by the reported experimental protocol, and the headline accuracy range is inconsistent with the authors' own Columbia results. As presented, the contribution is a limited empirical study whose main conclusions are not established; the significance depends on whether proper cross-dataset experiments can be added.

major comments (4)
  1. [Section 4.1, 4.3; Abstract] The claim of 'inter-dataset testing' and 'domain generalization' is not supported by the experimental design. The combined dataset D1+D2+D3 in Section 4.1 is built by pooling images from all three sources and then Section 4.3 reports results on a random train/test split of that pooled set. A random split of a pooled dataset allows the model to exploit source-specific cues (resolution, compression, sensor or generator artifacts) that are shared between training and test partitions, so high accuracy does not demonstrate generalization to an unseen dataset. To support the stated claim, the authors need a leave-one-dataset-out protocol (train on two datasets, test on the third) or an explicit train-on-one/test-on-other experiment. Without such experiments, the abstract's 'inter-dataset testing' and 'domain generalization' statements are unsupported.
  2. [Contributions list; Table 6] The contribution list states that the approach achieves 'state-of-the-art accuracy, consistently between 97-99% across multiple datasets.' This is contradicted by Table 6, where the Columbia dataset yields RGB accuracy of 0.95, F1-score of 0.80, and recall of 0.67, and the HSV space yields F1-score of 0.50. If 'multiple datasets' includes Columbia, the claim is false; if Columbia is excluded, the claim should be rephrased and justified. The discrepancy also appears in the conclusion, which acknowledges the low Columbia performance, making the headline claim internally inconsistent.
  3. [Section 4.2, Tables 4–8] All reported results appear to come from a single training run per configuration. There are no standard deviations, no repeated runs, and no statistical significance tests. This is especially problematic for the Columbia dataset, which contains only 43 CGI and 150 real images after curation (Section 4.1); the metrics in Table 6 vary dramatically across color spaces (recall 0.33–0.67), indicating that the results are likely unstable. The paper should report multiple runs (or cross-validation) with mean and variance, and should clearly describe how the small Columbia set was split for training and evaluation.
  4. [Section 4.2, Section 4.3, Table 8] The 'RGB is best' conclusion is presented as a finding, but RGB was selected after observing that it scored highest, and the paper then uses RGB for the comparative analysis. This is a post-hoc selection without a held-out validation of the color-space choice. The same data are used to select the color space and to report the headline accuracy, which is circular for the comparison across color spaces. The authors should either define a pre-registered hypothesis, use a validation set for color-space selection, or clearly label the per-color-space comparison as exploratory rather than confirmatory.
minor comments (6)
  1. [Section 4.3, JSSSTU subsection] The paragraph introducing JSSSTU results says 'The performance analysis of the CiFAKE dataset...' rather than referring to JSSSTU; this appears to be a copy-paste error.
  2. [Table 1, Table 2] Table 1 reports CiFAKE as having 1,20,000 images, while Table 2 reports '1,000,000+'; these numbers should be reconciled.
  3. [Throughout] The name of the Columbia dataset is spelled inconsistently ('columbia RCGI', 'Columbia PRCG', 'Columbia RCGI'); the canonical name should be used consistently.
  4. [Section 4.2] The paper does not describe how the combined dataset (D1+D2+D3) was exactly composed, e.g., how many images were taken from each dataset, whether the 1500-per-class count includes the augmented Columbia images, and what augmentation techniques were applied. This information is necessary for reproducibility.
  5. [Table 8 and Section 2] The 'state-of-the-art' claim in the contributions is not supported by a comparison with recent CGI detection methods; Table 8 compares only to VGG-19 and ResNet-50 on the combined RGB dataset, and no comparison is made on the individual datasets or against more recent transformer- or diffusion-aware detectors.
  6. [Figures 3–6] The t-SNE visualizations are described qualitatively as showing RGB 'most distinct separation,' but no quantitative measure of cluster separability is provided; consider reporting a silhouette score or similar metric.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical benchmark report; its color-space selection is post-hoc but transparent, and its inter-dataset claim is a validity concern rather than a derivation-level circularity.

full rationale

The paper makes no formal derivation claims that could reduce to their own inputs; it reports empirical evaluations of a Swin Transformer on three datasets and three color spaces. The sentence 'As a result, RGB was selected for domain generalization analysis' reflects post-hoc selection after comparing test accuracies, but the comparison across RGB, YCbCr, and HSV is fully reported in Tables 4-7 before the selection, so no fitted quantity is disguised as an independent prediction. The 'inter-dataset testing' terminology is imprecise: Section 4.1 constructs a pooled D1+D2+D3 set with 1500 images per class and Section 4.3 evaluates a random split of that pool, which is not a train-on-one-dataset/test-on-another domain-shift protocol; this is an evaluation-design limitation, not circularity. The abstract and contribution claim of '97-99% accuracy' also conflicts with Table 6's Columbia numbers (RGB recall 0.67, F1 0.80), but inconsistency is a correctness concern, not circularity. No load-bearing self-citations are present; the only self-reference is the acknowledgment of an earlier conference version, which is not used to justify any result. The reported accuracies are external benchmark measurements, so the paper is self-contained as an empirical study.

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

No invented entities. The paper introduces no free parameters beyond conventional training hyperparameters and learned network weights, which are standard and not fitted to a theory. The load-bearing assumptions are methodological: dataset labels are trusted, the damaged Columbia subset is treated as representative after unspecified augmentation, and the combined random split is interpreted as a domain-generalization test.

assumptions (5)
  • domain assumption Swin Transformer features are suitable for CGI and natural image discrimination.
    The method assumes a generic vision transformer trained with cross-entropy captures distinguishing traces of synthetic images; no forensic-specific design is added. This enters in Section 3.1.
  • domain assumption Dataset labels are correct ground truth.
    Accuracy is computed against CIFAKE, JSSSTU, and Columbia labels without a label audit. Entered in Section 4.1.
  • domain assumption The augmented Columbia subset is representative.
    Only 43 CGI and 150 real images were usable; augmentation is stated but not specified, and no evidence shows the expanded set preserves the original distribution. Entered in Section 4.1.
  • ad hoc to paper A random train/test split of the combined datasets measures domain generalization.
    The paper uses the mixed D1+D2+D3 split as its inter-dataset evaluation, but images from all sources appear in both training and test, so it does not test transfer to unseen sources. Entered in Sections 4.1 and 4.3.
  • standard math The standard Swin Transformer attention equations are accepted as background.
    Section 3.1 quotes the attention formula without proof or modification. This is background, not an invented component.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Swin Transformer for Robust CGI Images Detection: Intra- and Inter-Dataset Analysis across Multiple Color Spaces." pith.science (2026). https://pith.science/paper/L57YJFJK

@misc{pith2026250516253,
  author       = {Pith},
  title        = {Pith review of: Swin Transformer for Robust CGI Images Detection: Intra- and Inter-Dataset Analysis across Multiple Color Spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L57YJFJK}},
  note         = {Machine review of arXiv:2505.16253}
}
read the original abstract

This study aims to address the growing challenge of distinguishing computer-generated imagery (CGI) from authentic digital images across three different color spaces; RGB, YCbCr, and HSV. Given the limitations of existing classification methods in handling the complexity and variability of CGI, this research proposes a Swin Transformer based model for accurate differentiation between natural and synthetic images. The proposed model leverages the Swin Transformer's hierarchical architecture to capture local and global features for distinguishing CGI from natural images. Its performance was assessed through intra- and inter-dataset testing across three datasets: CiFAKE, JSSSTU, and Columbia. The model was evaluated individually on each dataset (D1, D2, D3) and on the combined datasets (D1+D2+D3) to test its robustness and domain generalization. To address dataset imbalance, data augmentation techniques were applied. Additionally, t-SNE visualization was used to demonstrate the feature separability achieved by the Swin Transformer across the selected color spaces. The model's performance was tested across all color schemes, with the RGB color scheme yielding the highest accuracy for each dataset. As a result, RGB was selected for domain generalization analysis and compared with other CNN-based models, VGG-19 and ResNet-50. The comparative results demonstrate the proposed model's effectiveness in detecting CGI, highlighting its robustness and reliability in both intra-dataset and inter-dataset evaluations. The findings of this study highlight the Swin Transformer model's potential as an advanced tool for digital image forensics, particularly in distinguishing CGI from natural images. The model's strong performance indicates its capability for domain generalization, making it a valuable asset in scenarios requiring precise and reliable image classification.

Figures

Figures reproduced from arXiv: 2505.16253 by the authors.

Figure 1
Figure 1. Example of few Computer Generated (CG) images from dataset CIFAKE-10, columbia RCGI and JSSSTU datasets (from top to bottom row), respectively. The images shows the difficulty in distingushing between two classes of images with naked eye. Also, the variation in computer generated images Traditional methods for addressing this problem have typically been divided into subjective and objective approaches. Subjective me… view at source ↗
Figure 2
Figure 2. Architecture of the proposed Swin transformer and the expansion of Swin Transformer Block under it. 3.1 Swin Transformer Architecture The Swin Transformer is a cutting-edge deep learning architecture renowned for effectively processing large-scale visual data with hierarchical representations. Unlike traditional convolutional neural networks (CNNs), the Swin Transformer adopts a hierarchical architecture that effici… view at source ↗
Figure 3
Figure 3. The plot illustrate the t-SNE plots of the extracted features from the Swin Transformer for the CiFake dataset (D1) images for the RGB, YCbCr and HSV color space (from left to right) (a) (b) (c) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The plot illustrate the t-SNE plots of the extracted features from the Swin Transformer for the JSSSTU dataset (D2) images for the RGB, YCbCr and HSV color space (from left to right) (a) (b) (c) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The plot illustrate the t-SNE plots of the extracted features from the Swin Transformer for the Columbia PRCG and real images dataset (D3) images for the RGB, YCbCr and HSV color space (from left to right) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The plot illustrate the t-SNE plots of the extracted features from the Swin Transformer for the combined datasets (D1+D2+D3) images for the RGB, YCbCr and HSV color space (from left to right) 4 Results and Discussion The Swin Transformer was trained utilizing features …
Figure 7
Figure 7. Figure 7: The training and validation accuracy plot, loss plot and ROC (from left to right) for the CiFAKE dataset for the RGB, YCbCr and HSV color frames (top to bottom). Performance on CiFAKE Dataset The performance analysis of the CiFAKE dataset for RGB, YCbCr, and HSV color …
Figure 8
Figure 8. Figure 8: The training and validation accuracy plot, loss plot and ROC (from left to right) for the JSSSTU dataset for the RGB, YCbCr and HSV color frames (top to bottom) [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: The training and validation accuracy plot, loss plot and ROC (from left to right) for the Columbia dataset for the RGB, YCbCr and HSV color frames (top to bottom) [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: The training and validation accuracy plot, loss plot and ROC (from left to right) for the Combined datasets (D1+D2+D3) for the RGB, YCbCr and HSV color frames (top to bottom) [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 24 canonical work pages

  1. [1]

    Cifake: Image classification and explainable identification of AI-generated synthetic images,

    J. J. Bird and A. Lotfi, "Cifake: Image classification and explainable identification of AI-generated synthetic images,"IEEE Access, 2024. 18 M. Preeti et al

    work page 2024

  2. [2]

    Columbia photographic images and photorealistic computer graphics dataset

    Tian-Tsong Ng, Shih-Fu Chang, Jessie Hsu, and Martin Pepeljugoski, "Columbia photographic images and photorealistic computer graphics dataset",Columbia Uni- versity, ADVENT Technical Report, pp. 205–2004, 2005

    work page 2004

  3. [3]

    Dataset for classification of computer graphic images and photographic images,

    Halaguru Basavarajappa Basanth Kumar and Haranahalli Rajanna Chennamma, “Dataset for classification of computer graphic images and photographic images,” IAES International Journal of Artificial Intelligence, vol. 11, no. 1, pp. 137, 2022

    work page 2022

  4. [4]

    Subjective and Objective Quality Assessment of Multi-Attribute Retouched Face Images,

    G. Yue, H. Wu, W. Yan, T. Zhou, H. Liu and W. Zhou, "Subjective and Objective Quality Assessment of Multi-Attribute Retouched Face Images," in IEEE Transactions on Broadcasting, vol. 70, no. 2, pp. 570-583, June 2024, doi: 10.1109/TBC.2024.3374043

    arXiv 2024

  5. [5]

    Exploring Rich Subjective Quality Information for Image Quality Assessment in the Wild

    Min, Xiongkuo, Yixuan Gao, Yuqin Cao, Guangtao Zhai, Wenjun Zhang, Huifang Sun, and Chang Wen Chen. "Exploring rich subjective quality information for image quality assessment in the wild." arXiv preprint arXiv:2409.05540, 2024

    work page Pith review arXiv 2024

  6. [6]

    How realistic is photorealistic?,

    S. Lyu and H. Farid, "How realistic is photorealistic?,"IEEE Transactions on Signal Processing, vol. 53, no. 2, pp. 845–850, 2005

    work page 2005

  7. [7]

    Identifying computer graphics using HSV color model and statistical moments of characteristic functions,

    W. Chen, Y. Q. Shi, and G. Xuan, "Identifying computer graphics using HSV color model and statistical moments of characteristic functions," in2007 IEEE Interna- tional Conference on Multimedia and Expo, 2007, pp. 1123–1126

    work page 2007

  8. [8]

    Physics-motivated features for distinguishing photographic images and computer graphics,

    T. Ng, S. Chang, J. Hsu, L. Xie, and M. Tsui, "Physics-motivated features for distinguishing photographic images and computer graphics," inProceedings of the 13th Annual ACM International Conference on Multimedia, 2005, pp. 239–248

    work page 2005

Show all 28 references
  1. [9]

    A Style-Based Generator Architecture for Generative Adversarial Net- works,

    Karras T., "A Style-Based Generator Architecture for Generative Adversarial Net- works," inarXiv preprintarXiv:1812.04948. 2019

    2019 arXiv

  2. [10]

    The deepfake detection challenge (dfdc) dataset,

    Dolhansky, Brian, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer, "The deepfake detection challenge (dfdc) dataset," arXiv preprint, arXiv:2006.07397, 2020

    2006 arXiv

  3. [11]

    Synthetic faces high quality (sfhq) dataset,

    Beniaguev, David, "Synthetic faces high quality (sfhq) dataset," 2022

    2022

  4. [12]

    Blind 3D-Synthesized Image Quality Measurement by Anal- ysis of Local and Global Statistical Properties,

    Z. Fang et al., "Blind 3D-Synthesized Image Quality Measurement by Anal- ysis of Local and Global Statistical Properties," in IEEE Transactions on In- strumentation and Measurement, vol. 72, pp. 1-15, 2023, Art no. 5024915, doi: 10.1109/TIM.2023.3306527

    2023

  5. [13]

    Large Scale GAN Training for High Fidelity Natural Image Syn- thesis,

    Brock, Andrew, "Large Scale GAN Training for High Fidelity Natural Image Syn- thesis," arXiv preprintarXiv:1809.11096, 2018

    2018 arXiv

  6. [14]

    An evaluation of deep learning- based computer generated image detection approaches,

    X. Ni, L. Chen, L. Yuan, G. Wu, and Y. Yao, "An evaluation of deep learning- based computer generated image detection approaches,"IEEE Access, vol. 7, pp. 130830–130840, 2019

    2019

  7. [15]

    Exposing computer generated images by eye’s region classification via trans- fer learning of VGG19,

    T. Carvalho, E. R. S. De Rezende, M. T. P. Alves, F. K. C. Balieiro, and R. B. Sovat, "Exposing computer generated images by eye’s region classification via trans- fer learning of VGG19," in2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA), 2...

    2017

  8. [16]

    Distinguishing between natural and computer-generated images using convolutional neural networks,

    W. Quan, K. Wang, D.-M. Yan, and X. Zhang, "Distinguishing between natural and computer-generated images using convolutional neural networks,"IEEE Trans- actions on Information Forensics and Security, vol. 13, no. 11, pp. 2772–2787, 2018

    2018

  9. [17]

    CGNet: Detecting computer-generated images based on transfer learning with attention module,

    Y. Yao, Z. Zhang, X. Ni, Z. Shen, L. Chen, and D. Xu, "CGNet: Detecting computer-generated images based on transfer learning with attention module,"Sig- nal Processing: Image Communication, vol. 105, p. 116692, 2022

    2022

  10. [18]

    Detecting differences between pho- tographs and computer generated images,

    J. Wu, M. V. Kamath, and S. Poehlman, “Detecting differences between pho- tographs and computer generated images,” inProceedings of the 24th IASTED in- ternational conference on Signal processing, pattern recognition, and applications, 2006, pp. 268–273. Title Suppressed Due t...

    2006

  11. [19]

    Digital image forensics for identifying com- puter generated and digital camera images,

    S. Dehnie, T. Sencar, and N. Memon, “Digital image forensics for identifying com- puter generated and digital camera images,” in2006 International Conference on Image Processing, 2006, pp. 2313–2316

    2006

  12. [20]

    Distinguishing computer graphics from photographic images using a multiresolution approach based on local binary patterns,

    Z. Li, Z. Zhang, and Y. Shi, “Distinguishing computer graphics from photographic images using a multiresolution approach based on local binary patterns,”Security and Communication Networks, vol. 7, no. 11, pp. 2153–2159, 2014

    2014

  13. [21]

    Identification of natural images and computer- generated graphics based on statistical and textural features,

    F. Peng, J.-t. Li, and M. Long, “Identification of natural images and computer- generated graphics based on statistical and textural features,”Journal of forensic sciences, vol. 60, no. 2, pp. 435–443, 2015

    2015

  14. [22]

    Distinguishing computer-generated images from pho- tographic images using two-stream convolutional neural network,

    K. B. Meena and V. Tyagi, "Distinguishing computer-generated images from pho- tographic images using two-stream convolutional neural network," Applied Soft Computing, vol. 100, pp. 107025, 2021

    2021

  15. [23]

    Locally GAN-generated face detection based on an improved Xception,

    B. Chen et al., "Locally GAN-generated face detection based on an improved Xception," Information Sciences, vol. 572, pp. 16–28, 2021

    2021

  16. [24]

    Detecting generated images by real images,

    B. Liu et al., "Detecting generated images by real images," inEuropean Conference on Computer Vision, 2022, pp. 95–110

    2022

  17. [25]

    Fake faces identification via convolutional neu- ral network,

    H. Mo, B. Chen, and W. Luo, "Fake faces identification via convolutional neu- ral network," inProceedings of the 6th ACM workshop on information hiding and multimedia security, 2018, pp. 43–47

    2018

  18. [26]

    Detecting Computer-Generated Images by Using Only Real Images

    Li, Ji, and Kai Wang, "Detecting Computer-Generated Images by Using Only Real Images." in17th International Conference on Machine Vision (ICMV 2024), 2024, pp. 1–11

    2024

  19. [27]

    Detection of AI-Generated Synthetic Images with a Lightweight CNN

    Ladjević, Adrian Lokner and Kramberger, Tin and Kramberger, Renata and Vla- hek, Dino, "Detection of AI-Generated Synthetic Images with a Lightweight CNN." AI 5.3 (2024): 1575

    2024

  20. [28]

    Detecting the Undetectable: Combining Kolmogorov-Arnold Networks and MLP for AI-Generated Image Detec- tion

    Anon, Taharim Rahman, and Jakaria Islam Emon. "Detecting the Undetectable: Combining Kolmogorov-Arnold Networks and MLP for AI-Generated Image Detec- tion." arXiv preprint arXiv:2408.09371 (2024)

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.