Pith. sign in

REVIEW 3 major objections 6 minor 71 references

Learning Counterfactually Decoupled Attention for Open-World Model Attribution

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that counterfactually separating attention into factual and counterfactual streams and maximizing their prediction gap improves open-world model attribution, especially for images from generative models never seen during…

desk verdict A useful plug-in for open-world model attribution with consistent empirical gains, but the causal decoupling mechanism has a sign error and an unsubstantiated link between loss and attention; worth refereeing after fixes. read the letter →

arxiv 2506.23074 v1 pith:UPNWQDVG submitted 2025-06-29 cs.CV cs.CRcs.LG

Yu Zheng , Boyang Gong , Fanye Kong , Yueqi Duan , Bingyao Yu , Wenzhao Zheng , Lei Chen , Jiwen Lu
show 1 more author
Jie Zhou
This is my paper
classification cs.CVcs.CRcs.LG
keywords open-worldmodelattributioncounterfactualattentioncausaleffectdeepfakeGANunseengeneralizationlearningsourcebiasdecoupling
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

Open-world model attribution asks which generative model produced an image, including models never seen during training. This paper claims that current methods, which rely on handcrafted region partitions or hand-designed feature spaces, are confounded by source-content biases (such as the identity of a face) and therefore fail on novel attacks. The proposed Counterfactually Decoupled Attention Learning (CDAL) trains two attention streams: a factual stream that highlights model-specific artifacts and a counterfactual stream that is pushed to highlight source-content bias, then maximizes the difference between their attribution predictions. On the OW-DFA and OSMA benchmarks the authors report that adding CDAL to existing methods improves attribution for unseen attacks by large margins at negligible computational cost. If the claimed causal decoupling is real, CDAL is a plug-and-play module rather than another task-specific detector.

What carries the argument

The machinery is the pair of attention maps $F$ and $C$: factual attention is extracted from the input feature map by Causal Expert (CE) convolutions, whose kernel is a dynamically weighted mixture of expert kernels, and counterfactual attention is extracted by a second CE branch. A decorrelation loss $L_{\text{decor}} = \operatorname{CE}(Y_c, Y_c)$ maximizes the entropy of counterfactual predictions, driving $C$ toward source-content regions that carry no attribution information. The causal-effect loss $L_{\text{causal}} = \operatorname{CE}(Y_f - Y_c, y)$ then trains the whole system so that factual attention outperforms counterfactual attention. A Causal Attention Augmentation step expands spatial coverage while preserving causal consistency by keeping factual regions intact and perturbing counterfactual regions.

What would settle it

Train CDAL on data where source content is artificially made predictive of the model (for example, each generative model is assigned a distinct image domain), then test on a held-out distribution of content. If the decoupling is causal, the counterfactual attention should track content and novel-attack gains should remain large; if the reported gains collapse when content is shuffled, or if the counterfactual branch highlights discriminative model-related regions, the causal-effect story is not what drives the improvements.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that the causal effect $Y_f - Y_c$ between factual and counterfactual attention predictions is a valid measure of how much a learned attention map captures generation-relevant traces, and that maximizing this effect with a cross-entropy loss teaches the network to ignore source-content biases and concentrate on patterns that transfer to unseen generative models. The authors show that their counterfactual attention, learned by Causal Expert Convolutions and an entropy-maximizing decorrelation loss rather than by randomized intervention, gives a better causal-effect estimate than static alternatives. They report that CDAL raises novel-attack attribution on OW-DFA by up to 11.27% in ARI over the CPL baseline and improves unseen-model purity on OSMA by up to 7.89%, with only about 0.35M additional parameters and 0.002 GFLOPs.

Load-bearing premise

The load-bearing premise is that the counterfactual branch, trained by entropy maximization and the causal-effect loss, genuinely isolates source-content bias rather than some other non-causal variation that happens to be uninformative about the known training models.

Editorial extensions

If this is right

  • Adding CDAL to an existing attribution model should keep its known-attack accuracy while substantially raising accuracy on novel, previously unseen attacks.
  • The method transfers across deepfake attribution and GAN attribution and discovery, so it is not tied to one generative family.
  • Because the causal-effect score quantifies attention quality, it can serve as a training signal without extra labels beyond the known-attack supervision.
  • The reported extension to diffusion and flow models implies the decoupling may keep working as new generative architectures appear.
  • With an efficiency overhead of roughly 0.35M parameters, CDAL can be grafted onto baselines without retraining them from scratch.

Reading between the lines

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

  • An unrun test that follows from the claim is to measure the mutual information between counterfactual attention maps and source-model identity: if the decoupling is real, this should be near zero, while factual attention maps should be highly informative.
  • The improvement on unseen seeds suggests the factual stream captures very subtle per-model fingerprints, raising the possibility of attributing at the level of training runs rather than architectures, which the current benchmarks do not explicitly require.
  • If the causal-effect gap is the true mechanism, then the factual branch alone should beat the baseline without the counterfactual head present at inference; the paper's ablations are consistent with this, but they do not ablate the branch at test time.
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 / 6 minor

Summary. The paper proposes Counterfactually Decoupled Attention Learning (CDAL), a plug-and-play module for open-world model attribution. CDAL extracts factual and counterfactual attention maps via causal expert convolutions, augments them through a causal attention augmentation strategy, and optimizes a causal-effect loss that maximizes the gap between factual and counterfactual predictions, together with decorrelation and augmentation losses. The method is evaluated on OW-DFA and OSMA benchmarks, integrated into several baselines (CPL, NACH, ORCA, RepMix, POSE), and reports consistent improvements, especially for unseen attacks, with small computational overhead.

Significance. If the causal-decoupling mechanism is validated, CDAL would be a practically valuable contribution: it is model-agnostic, lightweight, and improves open-set generalization across two independent benchmarks. The paper includes a thorough experimental comparison, component ablations, hyperparameter studies, efficiency analysis, and releases code. However, the central theoretical premise—that counterfactual attention trained with an entropy-based loss isolates source-content bias—is internally inconsistent as written and lacks direct verification; the significance is therefore conditional on resolving these issues.

major comments (3)
  1. [Sec. 3.2, Eq. (9)] The decorrelation loss is defined as L_decor = CE(Y_c, Y_c) = -Σ_c Y_c log Y_c, which is exactly the Shannon entropy H(Y_c). Equation (14) minimizes the total loss with a positive coefficient η_2 L_decor, so the implemented objective minimizes H(Y_c), pushing Y_c toward a one-hot distribution. The text in Sec. 3.2 states that 'By maximizing the entropy ... pushes Y_c towards a uniform distribution', which is the opposite of what Eq. (9) and Eq. (14) implement. If the released code actually maximizes L_decor, then Eq. (9) misrepresents the loss; if the code follows Eq. (9), the described mechanism is inverted. This contradiction must be resolved before the causal-decoupling claim can be accepted.
  2. [Sec. 3.2, Eqs. (6), (9), (12)] Even after correcting the sign, entropy maximization constrains only the prediction vector Y_c, not the attention map C(X). A degenerate counterfactual attention map that produces near-uniform class predictions would satisfy a maximized H(Y_c) without localizing any source-content bias. Conversely, nothing in the formulation ties C(X) to the semantic identity or content of the input. Consequently, the interpretation of Y_f - Y_c as the causal effect of removing model-specific artifacts (Sec. 3.1, Eq. 5) is not justified by the presented objectives. The authors need to provide direct evidence—for example, quantifying the overlap between C(X) and source-content regions, or an intervention study—that the learned counterfactual attention indeed isolates source biases rather than merely producing non-discriminative predictions.
  3. [Sec. 4, Tables 1-4 and Sec. 4.3] The main results are reported without error bars or standard deviations, and the hyperparameters η_1, η_2, η_3 and the number of experts N are selected based on ablation studies on the same benchmarks (Tables 5b and 5c). This creates a risk that the reported 'large margins' are partially due to tuning on the test benchmarks. Since the OSMA results are already averaged over five splits, standard deviations are available and should be reported. A held-out validation split for hyperparameter selection, or at least an explicit acknowledgment of this limitation, is needed to support the claim of consistent improvement.
minor comments (6)
  1. [Eq. (9)] The notation CE(Y_c, Y_c) is nonstandard; the expression actually equals the entropy H(Y_c), not a cross-entropy between two distinct distributions. Please clarify the notation or use H(Y_c) directly.
  2. [Table 4] For RepMix + Ours, the reported improvement in closed-set ACC is +0.32, but the values in the table (94.01 for RepMix vs. 93.83 for RepMix + Ours) imply a decrease of -0.18. The improvement row should be corrected.
  3. [Sec. 3.2] There are grammatical errors: 'our employ Causal Expert (CE) convolutions' should be 'we employ', and 'the denotes the estimated causal contribution factors' repeats 'the'.
  4. [Sec. 4.1] There are typos: 'dicrimination' should be 'discrimination', and 'effecive' should be 'effective'.
  5. [Tables 5a-5b] The column headers do not clearly align with the checkmarks; for example, it is ambiguous whether the first row in Table 5b includes L_causal or only the baseline. Please reformat the tables so that each column corresponds unambiguously to one loss component.
  6. [Throughout] The text inconsistently uses 'Eqn.' and 'Eq.' (e.g., 'Eqn. (12)' in Sec. 3.3 and 'Eq. (14)' in Sec. 3.2). Please standardize the equation citation style.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CDAL's causal-effect objective is a defined training loss evaluated on external benchmarks; the only self-citation [45] is non-load-bearing.

full rationale

The paper's central claim is that maximizing the causal effect Y_f - Y_c improves open-world model attribution. This is not a derivation of a prediction from fitted values: Y_f and Y_c are defined as outputs of factual and counterfactual attention branches (Eq. (5)), and Y_f - Y_c is then used as a supervised training target (Eq. (6)). The claim that this improves generalization is tested on external benchmarks (OW-DFA, OSMA, DF40-extended settings) rather than being read back from the training objective by construction. No fitted parameter is renamed as a prediction; no equation reduces to its own input. The self-citation to prior counterfactual attention learning ([45], Rao et al., with overlapping authors) appears only as related-work inspiration for the intervention idea, alongside independent citations to Pearl's causal framework [40,43], and it is not used to justify a uniqueness theorem or to forbid alternative designs. Therefore it is not load-bearing. The more substantive concern is internal inconsistency: Eq. (9) defines L_decor as CE(Y_c, Y_c) = -sum Y_c log Y_c, i.e., the Shannon entropy of the counterfactual prediction, while the text says it is maximized; however Eq. (14) minimizes the total loss with positive eta2, so the implemented objective would minimize entropy, contradicting the stated uniform-distribution goal. That is a correctness and reproducibility issue, not a circularity issue, because even a sign correction would not make the causal-decoupling claim equivalent to its inputs. Accordingly, no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim relies on the assumed causal graph, the validity of the counterfactual branch, and the definition of causal effect as a logit difference, plus benchmark protocols. No new physical entities are introduced; the CE convolution and augmentation are network modules, not free-standing entities.

free parameters (2)
  • eta_1, eta_2, eta_3 (loss weights) = not specified in main text
    Weights for L_causal, L_decor, L_aug in Eq. (14); tuned by hand or ablation per benchmark; central to loss balance.
  • N (number of causal expert kernels) = 4
    Selected via ablation (Table 5c); controls expressiveness of CE convolution.
assumptions (4)
  • domain assumption The SCM X -> A -> Y and X -> Y is the correct causal structure for attribution.
    Sec. 3.1 states this graph without testing alternatives; all subsequent causal-effect definitions depend on it.
  • ad hoc to paper Maximizing the entropy of counterfactual predictions forces C to focus on source content bias.
    Sec. 3.2; the paper asserts this decoupling but provides no quantitative verification beyond attention visualizations.
  • ad hoc to paper The difference of logits Y_f - Y_c measures the causal effect of model-specific artifacts.
    Sec. 3.1 Eq. (5)-(6); defined rather than derived from the SCM.
  • domain assumption The open-world benchmarks' labeled and unlabeled splits simulate novel attacks.
    Sec. 4.1-4.2; the paper relies on OW-DFA and OSMA protocols from prior work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Counterfactually Decoupled Attention for Open-World Model Attribution." pith.science (2026). https://pith.science/paper/UPNWQDVG

@misc{pith2026250623074,
  author       = {Pith},
  title        = {Pith review of: Learning Counterfactually Decoupled Attention for Open-World Model Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UPNWQDVG}},
  note         = {Machine review of arXiv:2506.23074}
}
read the original abstract

In this paper, we propose a Counterfactually Decoupled Attention Learning (CDAL) method for open-world model attribution. Existing methods rely on handcrafted design of region partitioning or feature space, which could be confounded by the spurious statistical correlations and struggle with novel attacks in open-world scenarios. To address this, CDAL explicitly models the causal relationships between the attentional visual traces and source model attribution, and counterfactually decouples the discriminative model-specific artifacts from confounding source biases for comparison. In this way, the resulting causal effect provides a quantification on the quality of learned attention maps, thus encouraging the network to capture essential generation patterns that generalize to unseen source models by maximizing the effect. Extensive experiments on existing open-world model attribution benchmarks show that with minimal computational overhead, our method consistently improves state-of-the-art models by large margins, particularly for unseen novel attacks. Source code: https://github.com/yzheng97/CDAL.

Figures

Figures reproduced from arXiv: 2506.23074 by the authors.

Figure 1
Figure 1. Upper: Problem Setup. Bottom: Exemplified sample in feature space. Bottom left: ArcFace [9] distances show that forgery images of the same source identity are clustered together regardless of their attacking models. Bottom middle: Feature dif￾ferences learned by existing methods [52] are still influenced by source bias (i.e., identity) from the novel attack rather than model￾specific artifacts. Bottom right: Our app… view at source ↗
Figure 2
Figure 2. Overview of our proposed CDAL, which can be readily incorporated into existing baseline networks. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of Casual Attention Augmentation. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of learned attention maps across various forgery types. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: t-sne visualization in OW-DFA. real faces simulating challenging real-world scenarios, our method also contributes to strong performance gains. Com￾pared to MPSL [53], CPL [52]+Ours leads by 1.37% in NMI for novel attacks. Integration into other baselines like ORCA [5]…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 63 canonical work pages

  1. [1]

    com / iperov / DeepFaceLab

    Deepfacelab.https : / / github . com / iperov / DeepFaceLab. Accessed: 2023-2-28. 6

    work page 2023

  2. [2]

    com / deepfakes / faceswap

    Deepfakes.https : / / github . com / deepfakes / faceswap. Accessed: 2023-2-28. 6

    work page 2023

  3. [3]

    Sit: Self-supervised vision transformer.arXiv preprint arXiv:2104.03602, 2021

    Sara Atito, Muhammad Awais, and Josef Kittler. Sit: Self-supervised vision transformer.arXiv preprint arXiv:2104.03602, 2021. 7

    arXiv 2021

  4. [4]

    Repmix: Represen- tation mixing for robust attribution of synthesized images

    Tu Bui, Ning Yu, and John Collomosse. Repmix: Represen- tation mixing for robust attribution of synthesized images. In ECCV, pages 146–163, 2022. 1, 2, 7

    work page 2022

  5. [5]

    Open-world semi-supervised learning

    Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. InICLR, 2022. 2, 6

    work page 2022

  6. [6]

    Stargan: Unified genera- tive adversarial networks for multi-domain image-to-image translation

    Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified genera- tive adversarial networks for multi-domain image-to-image translation. InCVPR, pages 8789–8797, 2018. 6

    work page 2018

  7. [7]

    Xception: Deep learning with depthwise separable convolutions

    Franc ¸ois Chollet. Xception: Deep learning with depthwise separable convolutions. InCVPR, pages 1251–1258, 2017. 4

    work page 2017

  8. [8]

    On the detection of digital face manipulation

    Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. InCVPR, pages 5781–5790, 2020. 6

    work page 2020

Show all 71 references
  1. [9]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InCVPR, pages 4690–4699, 2019. 1, 2

    2019

  2. [10]

    Reducing network agnostophobia

    Akshay Raj Dhamija, Manuel G ¨unther, and Terrance Boult. Reducing network agnostophobia. InNeurIPS, pages 9175– 9186, 2018. 7

    2018

  3. [11]

    Diffusion mod- els beat gans on image synthesis.NeurIPS, 34:8780–8794,

    Prafulla Dhariwal and Alexander Nichol. Diffusion mod- els beat gans on image synthesis.NeurIPS, 34:8780–8794,

  4. [12]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. InCVPR, pages 12873–12883, 2021. 1, 7

    2021

  5. [13]

    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, pages 3247–3258, 2020. 7

    2020

  6. [14]

    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, pages 14094– 14103, 2021. 2, 6, 7

    2021

  7. [15]

    Generative adversarial nets.NeurIPS, 27,

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.NeurIPS, 27,

  8. [16]

    On the exploitation of deepfake model recognition

    Luca Guarnera, Oliver Giudice, Matthias Nießner, and Se- bastiano Battiato. On the exploitation of deepfake model recognition. InCVPRW, pages 61–70, 2022. 1, 2

    2022

  9. [17]

    Robust semi-supervised learning when not all classes have labels

    Lan-Zhe Guo, Yi-Ge Zhang, Zhi-Fan Wu, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning when not all classes have labels. InNeurIPS, 2022. 2, 6

    2022

  10. [18]

    Automatically discov- ering and learning new visual categories with ranking statis- tics

    Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, An- drea Vedaldi, and Andrew Zisserman. Automatically discov- ering and learning new visual categories with ranking statis- tics. InICLR, 2020. 6

    2020

  11. [19]

    Ghostnet: More features from cheap opera- tions

    Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap opera- tions. InCVPR, pages 1580–1589, 2020. 4

    2020

  12. [20]

    Forgerynet: A versatile benchmark for comprehensive forgery analysis

    Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, and Ziwei Liu. Forgerynet: A versatile benchmark for comprehensive forgery analysis. InCVPR, pages 4360–4369, 2021. 6

    2021

  13. [21]

    Denoising dif- fusion probabilistic models.NeurIPS, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.NeurIPS, 33:6840–6851, 2020. 7

    2020

  14. [22]

    Pfa-gan: Progressive face aging with gen- erative adversarial network.TIFS, 16:2031–2045, 2020

    Zhizhong Huang, Shouzhen Chen, Junping Zhang, and Hongming Shan. Pfa-gan: Progressive face aging with gen- erative adversarial network.TIFS, 16:2031–2045, 2020. 1

    2020

  15. [23]

    Sc-fegan: Face editing gen- erative adversarial network with user’s sketch and color

    Youngjoo Jo and Jongyoul Park. Sc-fegan: Face editing gen- erative adversarial network with user’s sketch and color. In ICCV, pages 1745–1753, 2019. 6

    2019

  16. [24]

    Towards open world object de- tection

    KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vi- neeth N Balasubramanian. Towards open world object de- tection. InCVPR, pages 5830–5840, 2021. 2

    2021

  17. [25]

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

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. InICLR, 2018. 6

    2018

  18. [26]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401–4410, 2019. 1

    2019

  19. [27]

    Decentralized attribution of generative models

    Changhoon Kim, Yi Ren, and Yezhou Yang. Decentralized attribution of generative models. InICLR, 2021. 1, 2

    2021

  20. [28]

    Auto-encoding varia- tional bayes

    Diederik P Kingma and Max Welling. Auto-encoding varia- tional bayes. InICLR, 2014. 1

    2014

  21. [29]

    The hungarian method for the assignment problem.NRL, 2(1-2):83–97, 1955

    HW Kuhn et al. The hungarian method for the assignment problem.NRL, 2(1-2):83–97, 1955. 5

    1955

  22. [30]

    Are handcrafted filters helpful for attributing ai-generated images? InACM MM, pages 10698–10706, 2024

    Jialiang Li, Haoyue Wang, Sheng Li, Zhenxing Qian, Xin- peng Zhang, and Athanasios V Vasilakos. Are handcrafted filters helpful for attributing ai-generated images? InACM MM, pages 10698–10706, 2024. 1, 2

    2024

  23. [31]

    Faceshifter: Towards high fidelity and occlusion aware face swapping.arXiv preprint arXiv:1912.13457, 2019

    Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping.arXiv preprint arXiv:1912.13457, 2019. 6

    1912 arXiv

  24. [32]

    Face x-ray for more general face forgery detection

    Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. InCVPR, pages 5001–5010,

  25. [33]

    Counterfactual intervention feature transfer for visible- infrared person re-identification

    Xulin Li, Yan Lu, Bin Liu, Yating Liu, Guojun Yin, Qi Chu, Jinyang Huang, Feng Zhu, Rui Zhao, and Nenghai Yu. Counterfactual intervention feature transfer for visible- infrared person re-identification. InECCV, pages 381–398,

  26. [34]

    Celeb-df: A large-scale challenging dataset for deep- fake forensics

    Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deep- fake forensics. InCVPR, pages 3207–3216, 2020. 5

    2020

  27. [35]

    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, pages 95–110, 2022. 1

    2022

  28. [36]

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

    Fengyuan Liu, Haochen Luo, Yiming Li, Philip Torr, and Jindong Gu. Which model generated this image? a model- agnostic approach for origin attribution. InECCV, pages 282–301, 2024. 1, 2 9

    2024

  29. [37]

    Residual denoising diffu- sion models

    Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yan- dong Tang, and Liangqiong Qu. Residual denoising diffu- sion models. InCVPR, pages 2773–2783, 2024. 7

    2024

  30. [38]

    Discovering causal signals in images

    David Lopez-Paz, Robert Nishihara, Soumith Chintala, Bernhard Scholkopf, and L ´eon Bottou. Discovering causal signals in images. InCVPR, pages 6979–6987, 2017. 2

    2017

  31. [39]

    Do gans leave artificial fingerprints? In MIPR, pages 506–511, 2019

    Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In MIPR, pages 506–511, 2019. 1, 2, 7

    2019

  32. [40]

    Causality: models, reasoning, and inference, by judea pearl, cambridge university press, 2000

    Leland Gerson Neuberg. Causality: models, reasoning, and inference, by judea pearl, cambridge university press, 2000. Econometric Theory, 19(4):675–685, 2003. 2, 3, 4

    2000

  33. [41]

    FSGAN: Subject agnostic face swapping and reenactment

    Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. InICCV, pages 7184–7193, 2019. 6

    2019

  34. [42]

    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, pages 24480–24489, 2023. 1

    2023

  35. [43]

    Basic books, 2018

    Judea Pearl and Dana Mackenzie.The book of why: the new science of cause and effect. Basic books, 2018. 2, 3, 4

    2018

  36. [44]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InICCV, pages 4195–4205, 2023. 7

    2023

  37. [45]

    Counterfactual attention learning for fine-grained visual cat- egorization and re-identification

    Yongming Rao, Guangyi Chen, Jiwen Lu, and Jie Zhou. Counterfactual attention learning for fine-grained visual cat- egorization and re-identification. InICCV, pages 1025– 1034, 2021. 2

    2021

  38. [46]

    Openldn: Learn- ing to discover novel classes for open-world semi-supervised learning

    Mamshad Nayeem Rizve, Navid Kardan, Salman Khan, Fa- had Shahbaz Khan, and Mubarak Shah. Openldn: Learn- ing to discover novel classes for open-world semi-supervised learning. InECCV, pages 382–401, 2022. 2, 6

    2022

  39. [47]

    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, pages 10684– 10695, 2022. 1

    2022

  40. [48]

    Faceforen- sics++: Learning to detect manipulated facial images

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Chris- tian Riess, Justus Thies, and Matthias Nießner. Faceforen- sics++: Learning to detect manipulated facial images. In ICCV, pages 1–11, 2019. 5, 6

    2019

  41. [49]

    Open-world semantic segmen- tation including class similarity

    Matteo Sodano, Federico Magistri, Lucas Nunes, Jens Behley, and Cyrill Stachniss. Open-world semantic segmen- tation including class similarity. InCVPR, pages 3184–3194,

  42. [50]

    Learning structured output representation using deep conditional gen- erative models.NeurIPS, 28, 2015

    Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional gen- erative models.NeurIPS, 28, 2015. 1

    2015

  43. [51]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. InICLR, 2021. 1

    2021

  44. [52]

    Contrastive pseudo learning for open-world deepfake attribution

    Zhimin Sun, Shen Chen, Taiping Yao, Bangjie Yin, Ran Yi, Shouhong Ding, and Lizhuang Ma. Contrastive pseudo learning for open-world deepfake attribution. InICCV, pages 20882–20892, 2023. 1, 2, 3, 4, 5, 6, 7, 8

    2023

  45. [53]

    Rethinking open-world deepfake attribution with multi-perspective sensory learning.IJCV, pages 1–24, 2024

    Zhimin Sun, Shen Chen, Taiping Yao, Ran Yi, Shouhong Ding, and Lizhuang Ma. Rethinking open-world deepfake attribution with multi-perspective sensory learning.IJCV, pages 1–24, 2024. 1, 2, 3, 4, 5, 6

    2024

  46. [54]

    Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. InCVPR, pages 28130–28139, 2024. 1

    2024

  47. [55]

    Visual autoregressive modeling: Scalable image gen- eration via next-scale prediction.NeurIPS, 37:84839–84865,

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image gen- eration via next-scale prediction.NeurIPS, 37:84839–84865,

  48. [56]

    Neural discrete representation learning.NeurIPS, 30, 2017

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning.NeurIPS, 30, 2017. 1

    2017

  49. [57]

    Visualizing data using t-sne.JMLR, 9(11), 2008

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.JMLR, 9(11), 2008. 6

    2008

  50. [58]

    Counterfactual cycle-consistent learn- ing for instruction following and generation in vision- language navigation

    Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learn- ing for instruction following and generation in vision- language navigation. InCVPR, pages 15471–15481, 2022. 2, 4

    2022

  51. [59]

    Visual commonsense r-cnn

    Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense r-cnn. InCVPR, pages 10760– 10770, 2020. 2, 4

    2020

  52. [60]

    Forgerynir: deep face forgery and detec- tion in near-infrared scenario.TIFS, 17:500–515, 2022

    Yukai Wang, Chunlei Peng, Decheng Liu, Nannan Wang, and Xinbo Gao. Forgerynir: deep face forgery and detec- tion in near-infrared scenario.TIFS, 17:500–515, 2022. 6

    2022

  53. [61]

    Where did i come from? origin attribution of ai-generated images.NeurIPS, 36:74478–74500, 2023

    Zhenting Wang, Chen Chen, Yi Zeng, Lingjuan Lyu, and Shiqing Ma. Where did i come from? origin attribution of ai-generated images.NeurIPS, 36:74478–74500, 2023. 1, 2

    2023

  54. [62]

    Df40: Toward next- generation deepfake detection

    Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Chengjie Wang, Shouhong Ding, Yunsheng Wu, et al. Df40: Toward next- generation deepfake detection. InNeurIPS, 2024. 1, 6, 7

    2024

  55. [63]

    Deepfake network architecture attribution

    Tianyun Yang, Ziyao Huang, Juan Cao, Lei Li, and Xirong Li. Deepfake network architecture attribution. InAAAI, pages 4662–4670, 2022. 1, 2, 6, 7

    2022

  56. [64]

    Progressive open space expan- sion for open-set model attribution

    Tianyun Yang, Danding Wang, Fan Tang, Xinying Zhao, Juan Cao, and Sheng Tang. Progressive open space expan- sion for open-set model attribution. InCVPR, pages 15856– 15865, 2023. 1, 2, 3, 4, 5, 7, 8

    2023

  57. [65]

    Attributing fake images to gans: Learning and analyzing gan fingerprints

    Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In ICCV, pages 7556–7566, 2019. 1, 2, 7

    2019

  58. [66]

    Artificial fingerprinting for generative models: Root- ing deepfake attribution in training data

    Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Root- ing deepfake attribution in training data. InICCV, pages 14448–14457, 2021

    2021

  59. [67]

    Responsible disclosure of generative mod- els using scalable fingerprinting

    Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative mod- els using scalable fingerprinting. InICLR, 2021. 1, 2

    2021

  60. [68]

    Counterfactual zero-shot and open-set vi- sual recognition

    Zhongqi Yue, Tan Wang, Qianru Sun, Xian-Sheng Hua, and Hanwang Zhang. Counterfactual zero-shot and open-set vi- sual recognition. InCVPR, pages 15404–15414, 2021. 5

    2021

  61. [69]

    Multi-attentional deep- fake detection

    Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep- fake detection. InCVPR, pages 2185–2194, 2021. 1

    2021

  62. [70]

    Learning deep features for discrimi- native localization

    Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discrimi- native localization. InCVPR, pages 2921–2929, 2016. 5

    2016

  63. [71]

    Unpaired image-to-image translation using cycle- consistent adversarial networkss

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle- consistent adversarial networkss. InICCV, 2017. 6 10

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.