Computer Vision and Pattern Recognition · arXiv:2611.16307 v1 · submitted 12 Nov 2026 · 55 references
DEDCA: Test-Time Adaptation for Generalized AI-Generated Image Detection
The paper says
“AI-generated image detectors face significant challenges when deployed in real-world environments, particularly when test samples are produced by unseen generators that deviate from the source training distribution.”
more of the abstract
Existing detectors often rely on a single type of evidence, such as semantic representations or low-level generation traces, making them vulnerable when the corresponding cue becomes unreliable. Moreover, adapting detectors to newly emerging generators usually requires target labels or retraining, which is costly and impractical. To address these challenges, we propose Dual-Evidence Disagreement-Constrained Adaptation (DEDCA), a test-time adaptation framework for generalized AI-generated image detection. DEDCA constructs a dual-evidence detector by combining CLIP-based semantic evidence with trace evidence extracted from shuffled entropy and residual statistics, and introduces a reliability-aware gate to perform sample-specific adaptive fusion. Our key idea is to exploit the disagreement between the two evidence streams as an unlabeled adaptation signal, enabling the detector to adjust to target-domain shifts without relying on target labels or blindly trusting a single prediction stream. Experiments show that DEDCA achieves superior performance on the GenImage benchmark by outperforming state-of-the-art AI-generated image detectors and test-time adaptation methods with 92.0% average accuracy.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. DEDCA is a well-built detector with a marginal average win and an interesting but unvalidated disagreement-based TTA mechanism; worth reviewing, not worth trusting as-is. This line is wrong
5 major objections · 8 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 55 references say.
The paper says, at §3.4, Eq. (9); §4.3, Table 4
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.72):
“However, when the source detector is unreliable on unseen target data, such self-reinforcing objectives may amplify incorrect predictions and lead to confirmation bias or model collapse.”
Pith says: Objection 1 of 5 · major
The central novelty is the disagreement-constrained TTA loss Ltced (Eq. 9). The premise that cross-evidence disagreement reliably indicates target-domain miscalibration is never quantitatively validated on any held-out generator. The only support is the qualitative Fig. 1(b) and the leave-ADM ablation in Table 4.
more of objection 1
Without a direct correlation analysis (e.g., disagreement vs. error rate on target samples), the risk remains that minimizing Ltced could pull a correct branch toward an incorrect one, as noted in the stress-test. The authors acknowledge collapse risk in spirit but do not validate the premise. This is a load-bearing issue for the claimed novelty.
The paper says, at §4.2, Table 1
The review did not quote this passage.
Pith says: Objection 2 of 5 · major
The headline gain over MoLD is 92.02% vs 91.41% (p=0.021). While statistically significant, the margin is small (~0.6 absolute points). The practical significance of this improvement should be discussed, especially given the additional complexity of the dual-evidence design and TTA. The paper claims 'superior performance' but the margin is narrow against a single-stream detector.
The paper says, at §4.3, Table 4
The review did not quote this passage.
Pith says: Objection 3 of 5 · major
The ablation study for TTA is conducted only under the leave-ADM setting. While the authors note this is a challenging setting, the claim that Ltced provides the main adaptation gain (3.9 ACC drop when removed) is based on a single generator. Generalizability across other held-out generators is not demonstrated. The paper should either provide similar ablations for other generators or justify why ADM is representative.
The paper says, at §3.4, Eq. (10); §4.3, Table 4
The review did not quote this passage.
Pith says: Objection 4 of 5 · major
The anchoring losses Lgate and Ldrift are introduced to prevent drift, but their individual contributions are modest (removing Lgate drops ACC by 0.6, removing Ldrift by 1.6). The combined anchor loss (w/o Lanchor) drops ACC by 2.9, which is comparable to the Ltced drop of 3.9.
more of objection 4
This suggests that the anchoring mechanism is nearly as important as the disagreement signal itself. The paper should discuss whether the TTA gain is primarily from disagreement or from the combination with strong anchoring, and whether simpler TTA methods with similar anchoring would achieve comparable results.
The paper says, at §4.1, Implementation details
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.94):
“Further evaluation could also extend the current batch-transductive protocol to randomized target-stream orders and strictly prequential predict- then-adapt settings.”
Pith says: Objection 5 of 5 · major
The batch-transductive protocol (batch size 64, 10 adaptation steps per batch) is a specific setting. The paper acknowledges in the limitations (Appendix F) that further evaluation should extend to 'randomized target-stream orders and strictly prequential predict-then-adapt settings.' This is a significant limitation because the current protocol allows the model to see the entire batch before predicting, which is not strictly online. The impact of this protocol on the reported gains should be discussed.
8 minor objections
The paper says, at §1, Fig. 1
Pith says: minor objection 1 of 8
Figure 1(b) shows a qualitative correlation between disagreement and error probability, but the axes are not clearly labeled. The figure would benefit from a quantitative correlation coefficient or a scatter plot with actual data points.
The paper says, at §3.3, Eq. (5)
Pith says: minor objection 2 of 8
The cross-evidence disagreement loss Lced uses a stopgrad on the gate weight w. The rationale is clear, but the notation 'stopgrad' is non-standard. Consider using 'stop_gradient' or 'sg' for clarity.
show 6 more
The paper says, at §4.2, Table 1
Pith says: minor objection 3 of 8
The table reports average ACC but not standard deviations for all methods. The paper states DEDCA achieves 92.02±0.68% over five runs, but other methods' variability is not reported. This makes it difficult to assess the robustness of the comparison.
The paper says, at §4.3, Table 3 and Table 4
Pith says: minor objection 4 of 8
The ablation tables report performance drops as relative percentages (e.g., '↓11.8'). It would be clearer to also report the absolute ACC values for each variant, as the relative drop can be misleading when the baseline is low.
The paper says, at §3.1, SERT description
Pith says: minor objection 5 of 8
The SERT transform uses a fixed permutation generated with seed 42. The sensitivity of results to this permutation choice is not explored. Since the permutation is fixed across all experiments, it is not a source of variance, but the paper should note that the results may depend on this specific permutation.
The paper says, at Appendix C, Table 5
Pith says: minor objection 6 of 8
The robustness evaluation applies post-processing to both real and fake samples. This is a reasonable protocol, but the paper should clarify whether the adaptation is applied after post-processing or before. The timing of adaptation relative to post-processing affects the interpretation.
The paper says, at Appendix D, Table 6
Pith says: minor objection 7 of 8
The computational cost comparison shows DEDCA has 98.07M parameters and 2.84 ms inference time. The paper should discuss the memory and computational overhead of the dual-evidence design compared to simpler baselines like MoLD (303.71M params, 10.72 ms). DEDCA is more efficient than MoLD, but the comparison with MLEP (1.40M params) shows a large gap.
The paper says, at §5, Conclusion
Pith says: minor objection 8 of 8
The conclusion restates the contributions but does not discuss limitations or future work. The limitations are relegated to Appendix F, which is acceptable but should be referenced in the main conclusion.
What would settle it
Using the frozen source model M0 on each of the 8 held-out generators before any adaptation, compute for every target sample its cross-evidence disagreement JSD(ps,pt) and its fused-prediction error; report the per-generator Spearman correlation and the AUC of JSD as an error predictor (mirroring Fig. 1(b) but on actual target data). If the AUC is at or below chance (≈0.55) on the generators where the TTA gain is claimed (notably ADM), the disagreement-to-miscalibration premise fails and the Table 4 gain attributed to Ltced is most plausibly a generic entropy/regularization effect, leaving the central contribution unsubstantiated. A clearly positive correlation on all folds would resolve the concern.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 13 Nov 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 5 major objections, 8 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Short version: DEDCA is a competent, carefully engineered paper with a real but narrow empirical gain, and its central TTA premise remains asserted rather than demonstrated. I would send it to review, not reject it, but I would ask for substantial extra evidence before trusting the adaptation claim.
What's new: the system itself is a sensible combination — frozen CLIP semantic branch, shuffled multi-scale entropy plus residual maps for trace evidence, and a per-sample reliability gate that fuses the two logits. The gate-weighted JSD disagreement objective, used in both source training and test-time adaptation, is the genuine novel piece. Table 3 shows both branches matter and that removing L_ced hurts source training a lot; the effect sizes are large enough that the missing error bars on the ablations are a nuisance rather than disqualifying.
Now the soft spots, in order. The headline gain over MoLD is small: 92.02 versus 91.41, about 0.6 ACC points, with p=0.021 over five runs. Per generator, DEDCA wins five of eight and loses ADM and VQDM, so the average depends on balancing, not dominance. The TTA ablations are on one held-out generator (ADM), single runs, with no variance reported. More importantly, the load-bearing premise is unquantified: the paper claims cross-evidence disagreement indicates unreliable prediction (Fig. 1), but never reports a disagreement-to-error correlation on any held-out generator. This matters because Ltced is symmetric — minimizing JSD moves both branch probabilities toward their mean — and the gate weight only scales the penalty; nothing in the objective identifies which branch is wrong. The anchors (Lgate, Ldrift) plausibly prevent collapse, but they don't address the case where one branch is already correct and gets dragged toward the other. I would not call this a dealbreaker; it is an unvalidated mechanism. The authors do acknowledge the batch-transductive protocol and open-world limits in Section F, which I credit.
For a researcher in AI-generated image detection, this is a useful read: the implementation details are unusually complete, and the detector part is likely sound. The TTA contribution should be treated as promising but unproven. I would accept for peer review, with the request that per-generator error bars, repeated ablations, and a quantitative disagreement-versus-error study on held-out generators be added.
The strongest objection
DEDCA's central novelty is disagreement-constrained TTA: minimizing Ltced = w·JSD(ps,pt)^γ (Eq. 9) on unlabeled target batches is claimed to deliver the main adaptation gain. Table 4 attributes the dominant share of the leave-ADM improvement to Ltced (removing it drops ACC from 80.2 to 77.1, roughly 3.1 absolute points; the full TTA gain over no-TTA is 4.5 absolute points, reported by the authors as '5.6%', a relative figure). For this loss to adapt rather than collapse, two conditions must hold on unseen generators: (i) cross-evidence disagreement must reliably indicate target-domain miscalibration, and (ii) the symmetric update must not pull an already-correct branch toward an incorrect one. Neither is established. The only support offered is the qualitative Fig. 1(b) — 'samples with larger disagreement... are more likely to be misclassified' — plus the single-run, single-generator ablation in Table 4; no disagreement-to-error correlation is reported on any held-out generator. Structurally, Ltced couples ps and pt to their mean and updates both branches' biases; when exactly one branch is correct on a target sample, the joint gradient can corrupt the correct branch toward the wrong one. The gate weight w = 4m(1−m) only modulates the penalty's strength through stopgrad and does not bias which branch moves. The TTA component's entire claimed contribution, which is the paper's novelty, rests on this unquantified premise.
The reader's weakest_assumption — that cross-evidence disagreement is a reliable label-free indicator of target-domain miscalibration — is exactly the premise I judge most load-bearing, and I find no quantitative evidence in the manuscript that would retire it. The paper's stated contribution (abstract and §3.4) is the disagreement-constrained adaptation, and Table 4 shows Ltced is its largest single component: removing it costs roughly 3.1 absolute ACC points (80.2→77.1) of the 4.5-point total TTA gain on leave-ADM. That ablation is single-run, on one generator, without error bars, so it cannot distinguish the claimed mechanism from a generic disagreement/entropy regularizer whose benefit happens to appear on ADM.
Two technical sharpenings support treating this as the gate for the central claim. First, Ltced is symmetric: it couples ps and pt to their mean and updates both branches' biases, so where exactly one branch is already correct on a target sample, the joint gradient can pull the correct branch toward the wrong one. The weight w = 4m(1−m) gates the penalty by a stopgrad'd gate confidence and does not weight which branch moves more; whether this corruption occurs depends entirely on the unmeasured shape of disagreement-versus-error on target domains. Second, on the qualitative claim: Fig. 1(b) reports a monotone disagreement→error relationship, but no correlation coefficient or AUC is given, and the text does not establish that it was measured on held-out generators rather than illustrated.
I checked Table 4's internal consistency: the reported drops are relative percentages of the full model, e.g. (80.2−75.7)/80.2 ≈ 5.6%, so the text's '5.6% ACC' (b183) is a relative figure and the reader's '5.6 ACC points' mildly overstates the absolute gain of 4.5 points. This is a precision issue, not load-bearing. I also note that the batch-transductive protocol (adapt on a batch, then evaluate that same batch) is a real threat to deployment realism, but the authors explicitly acknowledge it in Section F (b374) as a future prequential extension; per my instructions I do not relabel that acknowledged limitation as a new finding, although the magnitude of its inflation on the reported gains remains unquantified.
Because the load-bearing premise is unresolved and the supporting ablation is thin, the reader's CONDITIONAL verdict stands; no adjustment is needed. The proposed test would settle the premise on actual target data, and its outcome is decisive for the core contribution.
The referee's report
If the central claim holds, DEDCA would advance the practical deployment of AI-generated image detectors by enabling adaptation to unseen generators without target labels. The dual-evidence design and disagreement-constrained TTA are conceptually interesting. The paper provides extensive experiments on a standard benchmark and includes ablation studies. However, the significance is tempered by the marginal gain over the strongest baseline (MoLD, 91.41% vs 92.02%, p=0.021) and the reliance on a single ablation setting (leave-ADM) for the core TTA claim.
Pith's reading of the paper (a machine's summary, not the authors')
AI-generated image detectors usually fail when asked to spot images made by a generator they never saw during training. DEDCA attacks this by using two independent sources of evidence at once, and by adapting on unlabeled test images rather than requiring fresh labels. The first source is semantic: a frozen CLIP vision-language encoder that captures high-level content. The second is trace evidence: the image is split into patches, scrambled into a fixed random order, converted to grayscale, and analyzed through local entropy maps, Laplacian high-pass residuals, and gradient magnitudes. Scrambling breaks object structure so the trace branch focuses on low-level generation artifacts. A small gating network, DERG, estimates per-sample reliability of each branch and fuses their logits. The key idea is that when the two branches disagree strongly on a test sample, at least one is likely mis-calibrated. DEDCA turns this disagreement into an unlabeled training signal: it penalizes large disagreement during adaptation, weighted by gate confidence, while entropy-minimizing the fused prediction and anchoring to the frozen source model to prevent drift. Only the gate, classifier biases, and trace BatchNorm affine parameters are updated; the CLIP encoder and trace convolutional weights stay frozen. On GenImage, holding out one generator at a time, DEDCA reports 92.0% average accuracy, about 0.6 points above the best prior method (MoLD) and about 2.6 points above its own detector without adaptation.
Core claim. DEDCA achieves 92.02±0.68% average ACC on the GenImage leave-one-generator benchmark, outperforming state-of-the-art detectors and TTA baselines, with the disagreement-constrained adaptation (Ltced, Eq. 9) supplying the main gain: 5.6 ACC points over the same detector without TTA on the leave-ADM setting (Table 4). The authors report both gains as significant under paired t-tests (p=0.021 vs MoLD; p<0.001 vs no-TTA).
The premise it rests on. Cross-evidence disagreement between the semantic and trace branches is a reliable, label-free indicator of target-domain miscalibration, so that minimizing JSD(ps,pt)^gamma (Eq. 9, weighted by gate confidence) safely adapts the detector to an unseen generator instead of driving both branches toward a shared, possibly wrong prediction. This premise is supported only by the qualitative motivation in Fig. 1 and the leave-ADM ablation in Table 4; no quantitative disagreement-to-error correlation on actual target domains is reported, and the anchors Lgate and Ldrif t (Eq. 10) are assumed sufficient to prevent collapse. The authors acknowledge the collapse risk in spirit but never validate the premise on a novel generator.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
AI-generated image detectors face significant challenges when deployed in real-world environments, particularly when test samples are produced by unseen generators that deviate from the source training distribution. Existing detectors often rely on a single type of evidence, such as semantic representations or low-level generation traces, making them vulnerable when the corresponding cue becomes unreliable. Moreover, adapting detectors to newly emerging generators usually requires target labels or retraining, which is costly and impractical. To address these challenges, we propose Dual-Evidence Disagreement-Constrained Adaptation (DEDCA), a test-time adaptation framework for generalized AI-generated image detection. DEDCA constructs a dual-evidence detector by combining CLIP-based semantic evidence with trace evidence extracted from shuffled entropy and residual statistics, and introduces a reliability-aware gate to perform sample-specific adaptive fusion. Our key idea is to exploit the disagreement between the two evidence streams as an unlabeled adaptation signal, enabling the detector to adjust to target-domain shifts without relying on target labels or blindly trusting a single prediction stream. Experiments show that DEDCA achieves superior performance on the GenImage benchmark by outperforming state-of-the-art AI-generated image detectors and test-time adaptation methods with 92.0% average accuracy.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-13
The authors' main claim, as the classifier picked it out 1.00
“Experiments show that DEDCA achieves superior performance on the GenImage benchmark by outperforming state-of-the-art AI-generated image detectors and test-time adaptation methods with 92.0\% average accuracy.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.09 · human participants 0.02 · controlled comparison 0.67 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.96 · mentions code or data availability 0.02
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| AI-generated image detectors face significant challenges when deployed in real-world environments, particularly when test samples are produced by unseen generators that deviate from the source training distribution. | background 0.99 | 0.06 | 0.02 | asserted 1.00 |
| Existing detectors often rely on a single type of evidence, such as semantic representations or low-level generation traces, making them vulnerable when the corresponding cue becomes unreliable. | background 0.60 | 0.04 | 0.02 | asserted 0.99 |
| Moreover, adapting detectors to newly emerging generators usually requires target labels or retraining, which is costly and impractical. | background 0.58 | 0.04 | 0.02 | asserted 1.00 |
| To address these challenges, we propose Dual-Evidence Disagreement-Constrained Adaptation (DEDCA), a test-time adaptation framework for generalized AI-generated image detection. | contribution 1.00 | 0.77 | 0.07 | asserted 0.98 |
| DEDCA constructs a dual-evidence detector by combining CLIP-based semantic evidence with trace evidence extracted from shuffled entropy and residual statistics, and introduces a reliability-aware gate to perform sample-specific adaptive fusion. | method 0.97 | 0.91 | 0.06 | asserted 1.00 |
| Our key idea is to exploit the disagreement between the two evidence streams as an unlabeled adaptation signal, enabling the detector to adjust to target-domain shifts without relying on target labels or blindly trusting a single prediction stream. | method 0.62 | 0.65 | 0.16 | asserted 0.99 |
| Experiments show that DEDCA achieves superior performance on the GenImage benchmark by outperforming state-of-the-art AI-generated image detectors and test-time adaptation methods with 92.0\% average accuracy. | result 1.00 | 0.97 | 0.74 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-13: 1 request, 5,712 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (7)
Figures are the authors', reproduced from the paper.
References (55)
- [1] Reza Babaei, Samuel Cheng, Rui Duan, and Shangqing Zhao. Generative artificial intelligence and the evolving challenge of deepfake detection: A systematic analysis.Journal of Sensor and Actuator Networks, 14(1):17, 2025 on Pith
- [2] Spiros Baxevanakis, Giorgos Kordopatis-Zilos, Panagiotis Galopoulos, Lazaros Apostolidis, Killian Levacher, Ipek Baris Schlicht, Denis Teyssou, Ioannis Kompatsiaris, and Symeon Papadopoulos. The mever deepfake detection service: Lessons learnt from developing and deploying in ... on Pith
- [3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis.arXiv preprint arXiv:1809.11096, 2018 on PithPith's reviewarXiv
- [4] Hanqun Cao, Cheng Tan, Zhangyang Gao, Yilun Xu, Guangyong Chen, Pheng-Ann Heng, and Stan Z Li. A survey on generative diffusion models.IEEE transactions on knowledge and data engineering, 36(7):2814–2830, 2024 on Pith
- [5] Qianying Cao, Somdatta Goswami, and George Em Karniadakis. Laplace neural operator for solving differential equations.Nature Machine Intelligence, 6(6):631–640, 2024 on Pith
- [6] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. InEuropean conference on computer vision, pages 103–120. Springer, 2020 on Pith
- [7] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdo- liva. On the detection of synthetic images generated by diffusion models. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), page... on Pith
- [8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021 on Pith
- [9] M Durgadevi et al. Generative adversarial network (gan): A general review on different variants of gan and applications. In2021 6th international conference on communication and electronics systems (ICCES), pages 1–8. IEEE, 2021. 10 on Pith
- [10] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696–10706, 2022 on Pith
- [11] Fabrizio Guillaro, Giada Zingarini, Ben Usman, Avneesh Sud, Davide Cozzolino, and Luisa Verdoliva. A bias-free training paradigm for more general ai-generated image detection. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 18685–18694, 2025 on Pith
- [12] Febryan Hidayat, Khasnan Hanif, and Maratus Sholikhah Miftakul Janah. The use of midjourney in design buying and selling business from the perspective of ibn ashur’s maqashid sharia.Indonesian Journal of Islamic Economic Law, 2(1):15–28, 2025 on Pith
- [13] Trung-Hieu Hoang, Duc M V o, et al. Persistent test-time adaptation in recurring testing scenarios.Advances in Neural Information Processing Systems, 37:123402–123442, 2024 on Pith
- [14] Zhenglin Huang, Jinwei Hu, Xiangtai Li, Yiwei He, Xingyu Zhao, Bei Peng, Baoyuan Wu, Xiaowei Huang, and Guangliang Cheng. Sida: Social media image deepfake detection, localization and explanation with large multimodal model. InProceedings of the Computer Vision and Pattern Rec... on Pith
- [15] Taehoon Kim, Jongwook Choi, Yonghyun Jeong, Haeun Noh, Jaejun Yoo, Seungryul Baek, and Jongwon Choi. Beyond spatial frequency: Pixel-wise temporal frequency-based deepfake video detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11198–... on Pith
- [16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014 on PithPith's reviewarXiv
- [17] Sangmin S Lee, Kiwon Lee, and Guiyeom Kang. Eeg artifact removal by bayesian deep learning & ica. In2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 932–935. IEEE, 2020 on Pith
- [18] Ziqiang Li, Jiazhen Yan, Ziwen He, Kai Zeng, Weiwei Jiang, Lizhi Xiong, and Zhangjie Fu. Is artificial intelligence generated image detection a solved problem?arXiv preprint arXiv:2505.12335, 2025 on PitharXiv
- [19] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts.International Journal of Computer Vision, 133(1):31–64, 2025 on Pith
- [20] Hanyang Lin, Yongzhao Zhan, Shiqin Liu, Xiao Ke, and Yuzhong Chen. A deep learning based bank card detection and recognition method in complex scenes: A deep learning based bank card detection and recognition method in complex scenes.Applied Intelligence, 52(13):15259–15277, 2022 on Pith
show 35 more
- [21] Huanqi Lin, Jinghui Qin, Xiaoqi Wu, Tianshui Chen, and Zhijing Yang. Revisiting dire: towards universal ai-generated image detection.Neural Networks, page 108084, 2025 on Pith
- [22] Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Jingdong Wang, and Yao Zhao. Forgery- aware adaptive transformer for generalizable synthetic image detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10770–10780, 2024 on Pith
- [23] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lareˆ 2: Latent reconstruction error based method for diffusion-generated image detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006–17015, 2024 on Pith
- [24] Hong-Hanh Nguyen-Le, Van-Tuan Tran, Dinh-Thuc Nguyen, and Nhien-An Le-Khac. Think twice before adaptation: Improving adaptability of deepfake detection via online test-time adaptation.arXiv preprint arXiv:2505.18787, 2025 on PithPith's reviewarXiv
- [25] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021 on PithPith's reviewarXiv
- [26] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. InInternational conference on machine learning, pages 16888–16905. PMLR, 2022 on Pith
- [27] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world.arXiv preprint arXiv:2302.12400, 2023 on PithPith's reviewarXiv
- [28] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24480–24489, 2023. 11 on Pith
- [29] NaHyeon Park, Kunhee Kim, Junsuk Choe, and Hyunjung Shim. Rethinking the use of vision transformers for ai-generated image detection.arXiv preprint arXiv:2512.04969, 2025 on PitharXiv
- [30] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. InEuropean conference on computer vision, pages 86–103. Springer, 2020 on Pith
- [31] Jinghui Qin, Jialin Chen, Tianshui Chen, and Zhijing Yang. Multimodal feature disentangle-fusion network for detecting and grounding multi-modal media manipulation.Expert Systems with Applications, page 132076, 2026 on Pith
- [32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag... on Pith
- [33] Md Shohel Rana, Mohammad Nur Nobi, Beddhu Murali, and Andrew H Sung. Deepfake detection: A systematic literature review.IEEE access, 10:25494–25513, 2022 on Pith
- [34] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130–9140, 2024 on Pith
- [35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022 on Pith
- [36] Chuangchuang Tan, Renshuai Tao, Huan Liu, Guanghua Gu, Baoyuan Wu, Yao Zhao, and Yunchao Wei. C2p-clip: Injecting category common prompt in clip to enhance generalization in deepfake detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages ... on Pith
- [37] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5052–5060, 2024 on Pith
- [38] 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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pag... on Pith
- [39] Francesco Tassone, Luca Maiano, and Irene Amerini. Continuous fake media detection: adapting deepfake detectors to new generative techniques.Computer Vision and Image Understanding, 249:104143, 2024 on Pith
- [40] Guillaume Vray, Devavrat Tomar, Xufeng Gao, Jean-Philippe Thiran, Evan Shelhamer, and Behzad Bozorgtabar. Reservoirtta: Prolonged test-time adaptation for evolving and recurring domains.arXiv preprint arXiv:2505.14511, 2025 on PitharXiv
- [41] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization.arXiv preprint arXiv:2006.10726, 2020 on PithPith's reviewarXiv
- [42] Hongsong Wang, Renxi Cheng, Yang Zhang, Chaolei Han, and Jie Gui. Lota: Bit-planes guided ai- generated image detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 17246–17255, 2025 on Pith
- [43] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7201–7211, 2022 on Pith
- [44] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695–8704, 2020 on Pith
- [45] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445–22455, 2023 on Pith
- [46] Shiyu Wu, Jing Liu, Jing Li, and Yequan Wang. Few-shot learner generalizes across ai-generated image detection.arXiv preprint arXiv:2501.08763, 2025. 12 on PithPith's reviewarXiv
- [47] Cheng Xia, Manxi Lin, Jiexiang Tan, Xiaoxiong Du, Yang Qiu, Junjun Zheng, Xiangheng Kong, Yun- ing Jiang, and Bo Zheng. Mirage: Towards ai-generated image detection in the wild.arXiv preprint arXiv:2508.13223, 2025 on PithPith's reviewarXiv
- [48] Shilin Yan, Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, and Weidi Xie. A sanity check for ai-generated image detection. InInternational Conference on Learning Representations, volume 2025, pages 70702–70720, 2025 on Pith
- [49] Lin Yuan, Xiaowan Li, Yan Zhang, Jiawei Zhang, Hongbo Li, and Xinbo Gao. Mlep: Multi-granularity local entropy patterns for universal ai-generated image detection.arXiv preprint arXiv:2504.13726, 2025 on PithPith's reviewarXiv
- [50] Hongji Zhang, Hao Tan, Jinghui Qin, and Zhijing Yang. Mff-net: Multi-view feature fusion network for ai-generated image detection.Expert Systems with Applications, page 134394, 2026 on Pith
- [51] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation.Advances in neural information processing systems, 35:38629–38642, 2022 on Pith
- [52] Qingyang Zhang, Yatao Bian, Xinke Kong, Peilin Zhao, and Changqing Zhang. Come: Test-time adaption by conservatively minimizing entropy.arXiv preprint arXiv:2410.10894, 2024 on PithPith's reviewarXiv
- [53] Chenming Zhou, Jiaan Wang, Yu Li, Lei Li, Juan Cao, and Sheng Tang. Beyond semantic features: Pixel-level mapping for generalized ai-generated image detection. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 36101–36109, 2026 on Pith
- [54] Ziyin Zhou, Yunpeng Luo, Yuanchen Wu, Ke Sun, Jiayi Ji, Ke Yan, Shouhong Ding, Xiaoshuai Sun, Yunsheng Wu, and Rongrong Ji. Aigi-holmes: Towards explainable and generalizable ai-generated image detection via multimodal large language models. InProceedings of the IEEE/CVF Inter... on Pith
- [55] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image.Advances in neural information processing systems, 36:77771–77782, 2023. A A Brief Intr... on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment