pith. machine review for the scientific record. sign in

arxiv: 2511.20068 · v2 · submitted 2025-11-25 · 💻 cs.CV

PRADA: Probability-Ratio-Based Attribution and Detection of Autoregressive-Generated Images

Simon Damm , Jonas Ricker , Henning Petzka , Asja Fischer This is my paper

Pith reviewed 2026-05-17 04:50 UTC · model grok-4.3

classification 💻 cs.CV
keywords autoregressive image generationsynthetic image detectionimage attributionprobability ratiotoken sequencedeepfake detectionAI-generated imagesconditional probability
0
0 comments X

The pith

The ratio of conditional to unconditional token probabilities uniquely marks images generated by a specific autoregressive model.

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

The paper introduces PRADA, a method that detects autoregressive-generated images and attributes them to their source model by examining the ratio of conditional and unconditional probabilities over the image's token sequence. This ratio produces patterns that differ systematically between images from one model, images from other models, and real photographs. The authors calibrate a simple per-model score from these ratios and apply thresholds to perform detection and attribution. Experiments demonstrate effectiveness on eight class-to-image and four text-to-image autoregressive generators. The approach is presented as interpretable because the signal derives directly from the model's own probability computations during generation.

Core claim

Whenever an image is generated by a particular autoregressive model, its probability ratio shows unique characteristics which are not present for images generated by other models or real images. These characteristics are exploited for threshold-based attribution and detection by calibrating a simple, model-specific score function based on the ratio of the model's conditional and unconditional probability for the autoregressive token sequence representing the given image.

What carries the argument

The ratio of a model's conditional probability to its unconditional probability for the autoregressive token sequence of the image.

If this is right

  • A single probability-ratio calculation per image enables attribution to one of several known autoregressive models without retraining complex classifiers.
  • The same ratio signal separates autoregressive-generated images from real images using model-specific thresholds.
  • The approach applies uniformly to both class-to-image and text-to-image autoregressive generators after per-model calibration.
  • Because the signal comes from the model's native probability estimates, it provides a direct link to the generation process itself.

Where Pith is reading between the lines

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

  • The same probability-ratio idea could be tested on sequential generators in other domains such as audio or video to see whether analogous signatures appear.
  • If the ratio patterns persist across model versions, they might serve as a lightweight fingerprint for tracing model families without full access to training data.
  • Pairing the ratio score with existing visual or statistical detectors could yield hybrid systems that remain effective when one cue is adversarially suppressed.

Load-bearing premise

The probability-ratio characteristics are sufficiently unique to each source model and stable enough across images to support reliable threshold-based detection and attribution even for unseen images and models.

What would settle it

If the calibrated score distributions for images from different autoregressive models overlap substantially or if scores for images from a new model fall into the range of a different model or of real images, the threshold-based attribution and detection would fail.

Figures

Figures reproduced from arXiv: 2511.20068 by Asja Fischer, Henning Petzka, Jonas Ricker, Simon Damm.

Figure 1
Figure 1. Figure 1: Overview of our proposed method. For any given image, PRADA extracts conditional and unconditional log-likelihoods for each token and assigns a score to their balanced ratio ∆α (xt, c). A lightweight calibration step provides a small, model-specific scoring function fθ : R → R. For next-scale prediction models, the scale-wise average of token scores is linearly combined with weights wi to obtain the final … view at source ↗
Figure 2
Figure 2. Figure 2: Distributions of features derived from log probabilities of AR image generators for real and generated images. While neither conditional probabilities (1st col), nor probability ratios (2nd col) or ICAS (3rd col) consistently tell real and generated images apart, our PRADA score separates their distributions. Learning a simple, token-wise scoring keeps our approach lightweight and interpretable. Scale Weig… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of the scale dependence. While higher scales are useful for VAR-d30, they are harmful for Infinity-2B, especially for ICAS. PRADA Score We can now combine all three compo￾nents (probability ratio balancing, token-wise scoring, and scale weighting) to obtain the PRADA score for an im￾age represented by tokens x. We assume that tokens are 2We find that weighting the means per scale rather than … view at source ↗
Figure 4
Figure 4. Figure 4: Attribution performance of PRADA. We report the confusion matrices (normalized over rows and averaged over five calibra￾tion runs) for class-to-image and text-to-image models. PRADA achieves high performance across various AR image generators and is particularly effective against text-to-image models. problem than detection [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Robustness analysis for PRADA. We report the AUROC for images generated by class-to-image (top) and text-to-image models (bottom) under varying degrees of perturbation. number of hidden neurons in each layer of fθ. We provide detailed results of our ablation study in Appendix E [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of image perturbations used to analyze the robustness of the proposed method. From top to bottom: JPEG compression, center crop, Gaussian blur and Gaussian noise at different strengths. 2 [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example images and PRADA scores for class-to-image models. ImageNet class labels (from left to right): 1 (goldfish), 28 (spotted salamander), 214 (Gordon setter), 604 (hourglass), 675 (moving van). 3 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Example images and PRADA scores for class-to-image models (continued). 4 [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example images and PRADA scores for text-to-image models. Generation prompts from Synthbuster [2] (from left to right): “statue of ‘Eros et Psyche’ in front of an ornamented glass window, in the style of museum, warm white balance”, “close-up of dark purple ´ tulips with large blooms high in the sun, soft-focus, 62mm f/4.8”, “a green grass glade surrounded by trees with lots of foliage”, “wrought iron brid… view at source ↗
Figure 10
Figure 10. Figure 10: Cumulative distributions of the log-probability ratio ∆(xt, c) over tokens for 200 random images and selected models. For all models, the log-probability ratio shows differences between real and generated images, but their differences are not uniform over different models. −4 −2 0 2 4 ∆(xt , c) −2.5 −2.0 −1.5 −1.0 −0.5 0.0 ICAS [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Token-wise scoring used by ICAS. In several scenarios, PRADA recovers a scoring function resembling the ICAS scor￾ing function. In this section, we visualize PRADA’s score function for different AR image generators and show how the learned parameters can help to understand the model’s probability distribution. We begin by reviewing why a uniform score function, such as ICAS, is not suitable for all models… view at source ↗
Figure 12
Figure 12. Figure 12: PRADA scoring for class-to-image models. The first column visualizes the differences in the cumulative distributions of α-balanced probability ratios ∆α (xt, c) over tokens for 100 random images, which are exploited by the scoring fθ(xt, c) in the second column. The third column visualizes fθ(xt, c) as function of conditional and unconditional likelihoods. The last column visualizes scale weights, showing… view at source ↗
Figure 13
Figure 13. Figure 13: PRADA scoring for class-to-image models (continued). stronger and more stable signal, and therefore a higher signal-to-noise ratio, than the earlier, coarser scales. As weight is still distributed across many scales, we also conclude that a single learnable function fθ can indeed work well across different scales. It remains to analyze a notable exception, namely Infinity-2B. Here, the finest scales recei… view at source ↗
Figure 14
Figure 14. Figure 14: PRADA scoring for text-to-image models. For additional context see the caption of [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Cancellation effect observed in the last two scales of Infinity-2B. We show heatmaps of ∆α (xt, c) for scale 12 vs. scale 13, for real and generates images (first 250 tokens per scale, over 1000 real and 1000 fake images). We observe that some tokens from real images cluster around ∆α (xt, c) ≈ 0 (left plot, top right corner). The negative scale weight w12 effectively eliminates the detrimental impacts ca… view at source ↗
Figure 16
Figure 16. Figure 16: Token-wise mean and standard deviation of ∆α (xt, c). Tokens are ordered from the coarsest to the finest scale and averaged over 10 000 real and 10 000 fake images for class-to-image models, and 1000 real and 1000 fake images for text-to-image models. We observe different behavior across scales, with coarser scales showing larger differences in the mean values, but also a larger standard deviation. This l… view at source ↗
Figure 17
Figure 17. Figure 17: Mean and standard deviation of ∆(xt, c) across scales. Comparing these to [PITH_FULL_IMAGE:figures/full_fig_p023_17.png] view at source ↗
read the original abstract

Autoregressive (AR) image generation has recently emerged as a powerful paradigm for image synthesis. Leveraging the generation principle of large language models, they allow for efficiently generating deceptively real-looking images, further increasing the need for reliable detection methods. However, to date there is a lack of work specifically targeting the detection of images generated by AR image generators. In this work, we present PRADA (Probability-Ratio-Based Attribution and Detection of Autoregressive-Generated Images), a simple and interpretable approach that can reliably detect AR-generated images and attribute them to their respective source model. The key idea is to inspect the ratio of a model's conditional and unconditional probability for the autoregressive token sequence representing a given image. Whenever an image is generated by a particular model, its probability ratio shows unique characteristics which are not present for images generated by other models or real images. We exploit these characteristics for threshold-based attribution and detection by calibrating a simple, model-specific score function. Our experimental evaluation shows that PRADA is highly effective against eight class-to-image and four text-to-image models. We release our code and data at github.com/jonasricker/prada.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces PRADA, a method for detecting and attributing images generated by autoregressive (AR) models. It computes the ratio of a model's conditional probability to its unconditional probability over the token sequence of a given image and uses model-specific thresholds on this ratio for detection and source attribution. The central claim is that these ratios exhibit unique, stable characteristics for images from a particular AR generator that are absent in real images or outputs from other models. Experiments are reported on eight class-to-image and four text-to-image AR models, with code and data released.

Significance. If the probability-ratio signatures prove unique and stable, the work offers a lightweight, interpretable detection approach for the emerging class of AR image generators, addressing a gap left by methods focused on GANs or diffusion models. The open release of code and data is a clear strength for reproducibility.

major comments (2)
  1. [Experimental Evaluation] The central claim that probability ratios produce non-overlapping, model-specific statistics rests on thresholds calibrated on a finite training set of images. No analysis is provided of how these distributions shift under changes in sampling temperature, prompt distribution, or architectural variants, which directly threatens the reliability of threshold-based attribution and detection for unseen images and models.
  2. [Method] The method defines a model-specific score function via simple thresholds on the probability ratio, yet the manuscript does not report sensitivity analysis or cross-validation details for threshold selection. This leaves open whether the reported effectiveness on the eight class-to-image and four text-to-image models generalizes beyond the calibration set.
minor comments (2)
  1. [Method] Notation for conditional versus unconditional probabilities should be introduced with an explicit equation early in the method section for clarity.
  2. [Figures] Figure captions could more explicitly state the number of images and models used in each panel to aid quick interpretation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the robustness of our threshold-based detection and attribution approach. We address each major comment below and describe planned revisions to improve the manuscript's rigor while preserving the core contributions of PRADA.

read point-by-point responses

  1. Referee: [Experimental Evaluation] The central claim that probability ratios produce non-overlapping, model-specific statistics rests on thresholds calibrated on a finite training set of images. No analysis is provided of how these distributions shift under changes in sampling temperature, prompt distribution, or architectural variants, which directly threatens the reliability of threshold-based attribution and detection for unseen images and models.

    Authors: We agree this is a substantive limitation in the current evaluation. While our experiments cover eight class-to-image and four text-to-image models with diverse architectures, we did not systematically vary sampling temperature, prompt distributions, or test architectural variants beyond those reported. The probability ratio is derived directly from each model's conditional and unconditional token probabilities, which we expect to reflect model-specific characteristics, but additional validation is warranted. In the revised manuscript we will add a sensitivity analysis section that includes experiments with multiple sampling temperatures on a subset of models and a discussion of how prompt variations affect the ratio distributions. We will also explicitly note the scope of generalization to unseen models as a limitation. revision: partial

  2. Referee: [Method] The method defines a model-specific score function via simple thresholds on the probability ratio, yet the manuscript does not report sensitivity analysis or cross-validation details for threshold selection. This leaves open whether the reported effectiveness on the eight class-to-image and four text-to-image models generalizes beyond the calibration set.

    Authors: Thresholds were selected on a calibration set of generated and real images to achieve high true-positive rates with low false positives on real data, as described in the experimental protocol. We did not include a formal sensitivity study or cross-validation procedure in the original submission. We will revise the method and experimental sections to provide explicit details on the calibration process, report performance across a range of threshold values, and include a sensitivity analysis showing how detection and attribution metrics vary with threshold choice. This will clarify the stability of the reported results. revision: yes

Circularity Check

0 steps flagged

No significant circularity in PRADA's probability-ratio method

full rationale

The paper computes the ratio of conditional to unconditional token probabilities directly from the AR model on a given image sequence and treats observed distributional differences versus real images or other generators as an empirical finding. Threshold calibration for detection/attribution is performed on held-out samples from the source models in a standard supervised manner. No step reduces the claimed uniqueness or detection performance to a self-definitional loop, a fitted parameter renamed as prediction, or a load-bearing self-citation. The derivation chain remains independent of its own outputs and is evaluated against external distributions (real images and alternate generators).

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The method relies on access to the internal conditional and unconditional probabilities of the target autoregressive models. Thresholds for detection and attribution are calibrated on held-out data, introducing model-specific parameters.

free parameters (1)
  • model-specific detection thresholds
    Calibrated per model to separate real, same-model, and other-model images based on the probability ratio distribution.

pith-pipeline@v0.9.0 · 5512 in / 1052 out tokens · 17615 ms · 2026-05-17T04:50:51.040859+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    The key idea is to inspect the ratio of a model's conditional and unconditional probability for the autoregressive token sequence... We exploit these characteristics for threshold-based attribution and detection by calibrating a simple, model-specific score function.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

86 extracted references · 86 canonical work pages

  1. [1]

    Source generator attribution via inversion

    Michael Albright and Scott McCloskey. Source generator attribution via inversion. InCVPRW, 2019. 3

    work page 2019

  2. [2]

    Synthbuster: Towards detection of diffu- sion model generated images.IEEE Open Journal of Signal Processing, 2023

    Quentin Bammey. Synthbuster: Towards detection of diffu- sion model generated images.IEEE Open Journal of Signal Processing, 2023. 6, 5

    work page 2023

  3. [3]

    A geometric and photo- metric exploration of GAN and diffusion synthesized faces

    Maty ´aˇs Boh ´aˇcek and Hany Farid. A geometric and photo- metric exploration of GAN and diffusion synthesized faces. InCVPRW, 2023. 2

    work page 2023

  4. [4]

    FakeInversion: Learning to detect images from un- seen text-to-image models by inverting Stable Diffusion

    George Cazenavette, Avneesh Sud, Thomas Leung, and Ben Usman. FakeInversion: Learning to detect images from un- seen text-to-image models by inverting Stable Diffusion. In CVPR, 2024. 2

    work page 2024

  5. [5]

    What makes fake images detectable? Understanding prop- erties that generalize

    Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? Understanding prop- erties that generalize. InECCV, 2020. 2

    work page 2020

  6. [6]

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

    Baoying Chen, Jishen Zeng, Jianquan Yang, and Rui Yang. DRCT: Diffusion reconstruction contrastive training towards universal detection of diffusion generated images. InICML,

    work page

  7. [7]

    Janus- pro: Unified multimodal understanding and generation with data and model scaling.arXiv Preprint, 2025

    Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Janus- pro: Unified multimodal understanding and generation with data and model scaling.arXiv Preprint, 2025. 3, 6, 1

    work page 2025

  8. [8]

    HFI: A unified framework for training-free detection and implicit water- marking of latent diffusion model generated images.arXiv Preprint, 2024

    Sungik Choi, Sungwoo Park, Jaehoon Lee, Seunghyun Kim, Stanley Jungkyu Choi, and Moontae Lee. HFI: A unified framework for training-free detection and implicit water- marking of latent diffusion model generated images.arXiv Preprint, 2024. 2

    work page 2024

  9. [9]

    FIRE: Robust detection of diffusion- generated images via frequency-guided reconstruction error

    Beilin Chu, Xuan Xu, Xin Wang, Yufei Zhang, Weike You, and Linna Zhou. FIRE: Robust detection of diffusion- generated images via frequency-guided reconstruction error. InCVPR, 2025. 2

    work page 2025

  10. [10]

    Word association norms, mutual information, and lexicography.Computa- tional linguistics, 1990

    Kenneth Church and Patrick Hanks. Word association norms, mutual information, and lexicography.Computa- tional linguistics, 1990. 4

    work page 1990

  11. [11]

    Fast and accurate deep network learning by exponential linear units (elus)

    Djork-Arn ´e Clevert, Thomas Unterthiner, and Sepp Hochre- iter. Fast and accurate deep network learning by exponential linear units (elus). InICLR, 2016. 4

    work page 2016

  12. [12]

    Intriguing properties of syn- thetic images: From generative adversarial networks to dif- fusion models

    Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of syn- thetic images: From generative adversarial networks to dif- fusion models. InCVPRW, 2023. 2

    work page 2023

  13. [13]

    Towards universal GAN image detection

    Davide Cozzolino, Diego Gragnaniello, Giovanni Poggi, and Luisa Verdoliva. Towards universal GAN image detection. In VCIP, 2021. 2

    work page 2021

  14. [14]

    Raising the bar of AI-generated image detection with CLIP

    Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of AI-generated image detection with CLIP. InCVPRW, 2024. 2

    work page 2024

  15. [15]

    Zero-shot detection of AI-generated im- ages

    Davide Cozzolino, Giovanni Poggi, Matthias Nießner, and Luisa Verdoliva. Zero-shot detection of AI-generated im- ages. InECCV, 2024. 2

    work page 2024

  16. [16]

    RAISE: A raw images dataset for dig- ital image forensics

    Duc-Tien Dang-Nguyen, Cecilia Pasquini, Valentina Conot- ter, and Giulia Boato. RAISE: A raw images dataset for dig- ital image forensics. InMMSYS, 2015. 6

    work page 2015

  17. [17]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. InCVPR, 2009. 6

    work page 2009

  18. [18]

    Diffusion models beat GANs on image synthesis.NeurIPS, 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis.NeurIPS, 2021. 3

    work page 2021

  19. [19]

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

    Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: CNN based generative deep neural networks are failing to reproduce spectral distributions. In CVPR, 2020. 2

    work page 2020

  20. [20]

    Fourier spectrum discrepancies in deep network generated images

    Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. InNeurIPS, 2020. 2

    work page 2020

  21. [21]

    Lighting (in)consistency of paint by text.arXiv Preprint, 2022

    Hany Farid. Lighting (in)consistency of paint by text.arXiv Preprint, 2022. 2

    work page 2022

  22. [22]

    Perspective (in)consistency of paint by text

    Hany Farid. Perspective (in)consistency of paint by text. arXiv Preprint, 2022. 2

    work page 2022

  23. [23]

    Criminals use generative artificial intelligence to facilitate financial fraud.https: //www.ic3.gov/PSA/2024/PSA241203, 2024

    Federal Bureau of Investigation. Criminals use generative artificial intelligence to facilitate financial fraud.https: //www.ic3.gov/PSA/2024/PSA241203, 2024. 1

    work page 2024

  24. [24]

    Leveraging fre- quency analysis for deep fake image recognition

    Joel Frank, Thorsten Eisenhofer, Lea Sch ¨onherr, Asja Fis- cher, Dorothea Kolossa, and Thorsten Holz. Leveraging fre- quency analysis for deep fake image recognition. InICML,

    work page

  25. [25]

    A representative study on human detection of artificially generated media across countries

    Joel Frank, Franziska Herbert, Jonas Ricker, Lea Sch ¨onherr, Thorsten Eisenhofer, Asja Fischer, Markus D ¨urmuth, and Thorsten Holz. A representative study on human detection of artificially generated media across countries. InIEEE Sym- posium on Security and Privacy (S&P), 2024. 1

    work page 2024

  26. [26]

    Towards discovery and attribution of open-world GAN generated images

    Sharath Girish, Saksham Suri, Sai Saketh Rambhatla, and Abhinav Shrivastava. Towards discovery and attribution of open-world GAN generated images. InICCV, 2021. 3

    work page 2021

  27. [27]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. InNeurIPS,

    work page

  28. [28]

    Gragnaniello, D

    D. Gragnaniello, D. Cozzolino, F. Marra, G. Poggi, and L. Verdoliva. Are GAN generated images easy to detect? A critical analysis of the state-of-the-art. InICME, 2021. 2

    work page 2021

  29. [29]

    A bias-free training paradigm for more general AI-generated image de- tection

    Fabrizio Guillaro, Giada Zingarini, Ben Usman, Avneesh Sud, Davide Cozzolino, and Luisa Verdoliva. A bias-free training paradigm for more general AI-generated image de- tection. InCVPR, 2025. 2

    work page 2025

  30. [30]

    Eyes tell all: Irregular pupil shapes reveal GAN- generated faces

    Hui Guo, Shu Hu, Xin Wang, Ming-Ching Chang, and Si- wei Lyu. Eyes tell all: Irregular pupil shapes reveal GAN- generated faces. InICASSP, 2022. 2

    work page 2022

  31. [31]

    Infinity: Scaling bit- wise autoregressive modeling for high-resolution image syn- thesis

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bit- wise autoregressive modeling for high-resolution image syn- thesis. InCVPR, 2025. 3, 4, 6, 1

    work page 2025

  32. [32]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR,

    work page

  33. [33]

    RIGID: A training-free and model-agnostic framework for robust ai- generated image detection.arXiv Preprint, 2024

    Zhiyuan He, Pin-Yu Chen, and Tsung-Yi Ho. RIGID: A training-free and model-agnostic framework for robust ai- generated image detection.arXiv Preprint, 2024. 2, 6, 1

    work page 2024

  34. [34]

    Did you use my GAN to generate fake? Post-hoc 9 attribution of GAN generated images via latent recovery

    Syou Hirofumi, Kazuto Fukuchi, Yohei Akimoto, and Jun Sakuma. Did you use my GAN to generate fake? Post-hoc 9 attribution of GAN generated images via latent recovery. In IJCNN, 2022. 3

    work page 2022

  35. [35]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3

    work page 2021

  36. [36]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. InNeurIPS, 2020. 1

    work page 2020

  37. [37]

    Can we no longer be- lieve anything we see?The New York Times, 2023.https: //www.nytimes.com/2023/04/08/business/ media/ai-generated-images.html

    Tiffany Hsu and Steven Lee Myers. Can we no longer be- lieve anything we see?The New York Times, 2023.https: //www.nytimes.com/2023/04/08/business/ media/ai-generated-images.html. 1

    work page 2023

  38. [38]

    Exposing GAN- Generated faces using inconsistent corneal specular high- lights

    Shu Hu, Yuezun Li, and Siwei Lyu. Exposing GAN- Generated faces using inconsistent corneal specular high- lights. InICASSP, 2021. 2

    work page 2021

  39. [39]

    Generalized image- based deepfake detection through foundation model adapta- tion

    Tai-Ming Huang, Yue-Hua Han, Ernie Chu, Shu-Tzu Lo, Kai-Lung Hua, and Jun-Cheng Chen. Generalized image- based deepfake detection through foundation model adapta- tion. InICPR, 2025. 2

    work page 2025

  40. [40]

    Leveraging rep- resentations from intermediate encoder-blocks for synthetic image detection

    Christos Koutlis and Symeon Papadopoulos. Leveraging rep- resentations from intermediate encoder-blocks for synthetic image detection. InECCV, 2025. 2

    work page 2025

  41. [41]

    Privacy attacks on image AutoRegressive models

    Antoni Kowalczuk, Jan Dubi ´nski, Franziska Boenisch, and Adam Dziedzic. Privacy attacks on image AutoRegressive models. InICML, 2025. 1, 4, 8

    work page 2025

  42. [42]

    HMAR: Efficient hierarchical masked autoregressive image generation

    Hermann Kumbong, Xian Liu, Tsung-Yi Lin, Xihui Liu, Zi- wei Liu, Daniel Y Fu, Ming-Yu Liu, Christopher Re, and David W Romero. HMAR: Efficient hierarchical masked autoregressive image generation. InCVPR, 2025. 3, 6, 1

    work page 2025

  43. [43]

    Single-model attribution of generative models through final-layer inversion

    Mike Laszkiewicz, Jonas Ricker, Johannes Lederer, and Asja Fischer. Single-model attribution of generative models through final-layer inversion. InICML, 2024. 3

    work page 2024

  44. [44]

    BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InICML,

    work page

  45. [45]

    Detecting generated images by real images

    Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In ECCV, 2022. 2

    work page 2022

  46. [46]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InICLR, 2019. 5

    work page 2019

  47. [47]

    LaREˆ2: Latent reconstruction error based method for diffusion-generated image detection

    Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. LaREˆ2: Latent reconstruction error based method for diffusion-generated image detection. InCVPR, 2024. 2

    work page 2024

  48. [48]

    Do GANs leave artificial fingerprints? In MIPR, 2019

    Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do GANs leave artificial fingerprints? In MIPR, 2019. 3

    work page 2019

  49. [49]

    Ex- ploiting visual artifacts to expose deepfakes and face manip- ulations

    Falko Matern, Christian Riess, and Marc Stamminger. Ex- ploiting visual artifacts to expose deepfakes and face manip- ulations. InWACVW, 2019. 2

    work page 2019

  50. [50]

    Detecting GAN- generated imagery using saturation cues

    Scott McCloskey and Michael Albright. Detecting GAN- generated imagery using saturation cues. InICIP, 2019. 2

    work page 2019

  51. [51]

    Do deep generative models know what they don’t know? InICLR, 2019

    E Nalisnick, A Matsukawa, Y Teh, D Gorur, and B Laksh- minarayanan. Do deep generative models know what they don’t know? InICLR, 2019. 1

    work page 2019

  52. [52]

    Lakshmanan Nataraj, Tajuddin Manhar Mohammed, B. S. Manjunath, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H. Bappy, and Amit K. Roy-Chowdhury. Detecting GAN generated fake images using co-occurrence matrices. Electronic Imaging, 2019. 2

    work page 2019

  53. [53]

    Towards uni- versal fake image detectors that generalize across generative models

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards uni- versal fake image detectors that generalize across generative models. InCVPR, 2023. 2

    work page 2023

  54. [54]

    The creativity of text-to-image gener- ation

    Jonas Oppenlaender. The creativity of text-to-image gener- ation. InProceedings of the 25th International Academic Mindtrek Conference, 2022. 1

    work page 2022

  55. [55]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InICML, 2021. 3

    work page 2021

  56. [56]

    On the effectiveness of dataset alignment for fake image detection

    Anirudh Sundara Rajan, Utkarsh Ojha, Jedidiah Schloesser, and Yong Jae Lee. On the effectiveness of dataset alignment for fake image detection. InICLR, 2025. 2

    work page 2025

  57. [57]

    Hierarchical text-conditional image gener- ation with CLIP latents.arXiv Preprint, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with CLIP latents.arXiv Preprint, 2022. 1

    work page 2022

  58. [58]

    Likelihood ratios for out-of-distribution detec- tion.NeurIPS, 2019

    Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshmi- narayanan. Likelihood ratios for out-of-distribution detec- tion.NeurIPS, 2019. 1

    work page 2019

  59. [59]

    Towards the detection of diffusion model deepfakes

    Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fis- cher. Towards the detection of diffusion model deepfakes. In VISIGRAPP, 2024. 2

    work page 2024

  60. [60]

    AER- OBLADE: Training-free detection of latent diffusion images using autoencoder reconstruction error

    Jonas Ricker, Denis Lukovnikov, and Asja Fischer. AER- OBLADE: Training-free detection of latent diffusion images using autoencoder reconstruction error. InCVPR, 2024. 2, 6

    work page 2024

  61. [61]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InCVPR, 2022. 1

    work page 2022

  62. [62]

    How generative AI is boosting the spread of disinformation and propaganda.MIT Technology Re- view, 2023.https : / / www

    Tate Ryan-Mosley. How generative AI is boosting the spread of disinformation and propaganda.MIT Technology Re- view, 2023.https : / / www . technologyreview . com / 2023 / 10 / 04 / 1080801 / generative - ai - boosting- disinformation- and- propaganda- freedom-house/. 1

    work page 2023

  63. [63]

    Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1

    work page 2022

  64. [64]

    Ayush Sarkar, Hanlin Mai, Amitabh Mahapatra, Svetlana Lazebnik, D. A. Forsyth, and Anand Bhattad. Shadows don’t lie and lines can’t bend! Generative models don’t know pro- jective geometry...for now. InCVPR, 2024. 2

    work page 2024

  65. [65]

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

    Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. DE- FAKE: Detection and attribution of fake images generated by text-to-image generation models. InCCS, 2023. 2

    work page 2023

  66. [66]

    Detecting pretraining data from large language mod- els

    Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettle- moyer. Detecting pretraining data from large language mod- els. InICLR, 2024. 4 10

    work page 2024

  67. [67]

    Generative modeling by es- timating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by es- timating gradients of the data distribution. InNeurIPS, 2019. 1

    work page 2019

  68. [68]

    Autoregressive model beats diffusion: Llama for scalable image generation.arXiv Preprint, 2024

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation.arXiv Preprint, 2024. 1, 3, 6

    work page 2024

  69. [69]

    Synthetic image verification in the era of generative artificial intelli- gence: What works and what isn’t there yet.IEEE Security & Privacy, 2024

    Diangarti Tariang, Riccardo Corvi, Davide Cozzolino, Gio- vanni Poggi, Koki Nagano, and Luisa Verdoliva. Synthetic image verification in the era of generative artificial intelli- gence: What works and what isn’t there yet.IEEE Security & Privacy, 2024. 2

    work page 2024

  70. [70]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Li- wei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. InNeurIPS, 2024. 1, 3, 4, 6

    work page 2024

  71. [71]

    Pixel recurrent neural networks

    A ¨aron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. InICML,

    work page

  72. [72]

    Neural discrete representation learning

    Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. InNeurIPS, 2017. 3

    work page 2017

  73. [73]

    Switti: Design- ing scale-wise transformers for text-to-image synthesis

    Anton V oronov, Denis Kuznedelev, Mikhail Khoroshikh, Valentin Khrulkov, and Dmitry Baranchuk. Switti: Design- ing scale-wise transformers for text-to-image synthesis. In CVPR, 2025. 3, 6, 1

    work page 2025

  74. [74]

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. CNN-generated images are surprisingly easy to spot... for now. InCVPR, 2020. 2

    work page 2020

  75. [75]

    DIRE for diffusion-generated image detection

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. DIRE for diffusion-generated image detection. InICCV, 2023. 2

    work page 2023

  76. [76]

    Where did I come from? Origin attribution of AI-generated images.NeurIPS, 2023

    Zhenting Wang, Chen Chen, Yi Zeng, Lingjuan Lyu, and Shiqing Ma. Where did I come from? Origin attribution of AI-generated images.NeurIPS, 2023. 3

    work page 2023

  77. [77]

    Metaxas, and Shiqing Ma

    Zhenting Wang, Vikash Sehwag, Chen Chen, Lingjuan Lyu, Dimitris N. Metaxas, and Shiqing Ma. How to trace latent generative model generated images without artificial water- mark? InICML, 2024. 3

    work page 2024

  78. [78]

    Janus: Decoupling visual encoding for unified multimodal understanding and genera- tion.arXiv Preprint, 2024

    Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, and Ping Luo. Janus: Decoupling visual encoding for unified multimodal understanding and genera- tion.arXiv Preprint, 2024. 3

    work page 2024

  79. [79]

    Learning to disentangle GAN fingerprint for fake image attribution.arXiv Preprint, 2021

    Tianyun Yang, Juan Cao, Qiang Sheng, Lei Li, Jiaqi Ji, Xirong Li, and Sheng Tang. Learning to disentangle GAN fingerprint for fake image attribution.arXiv Preprint, 2021. 3

    work page 2021

  80. [80]

    Icas: Detecting training data from autoregressive image gen- erative models

    Hongyao Yu, Yixiang Qiu, Yiheng Yang, Hao Fang, Tianqu Zhuang, Jiaxin Hong, Bin Chen, Hao Wu, and Shu-Tao Xia. Icas: Detecting training data from autoregressive image gen- erative models. InACM MM, 2025. 1, 4, 8, 7

    work page 2025

Showing first 80 references.