acceptodds
Under review as a conference paper at ICLR 2027

EviCal: Learning and Calibrating Evidence for Generalizable AI-Generated Image Detection

13 views

Abstract

AI-generated image detectors can perform well within a benchmark yet degrade when generators and real-image sources change. Existing detectors may learn artifacts specific to their training generators, while their decision rule may mistake normal variation in unfamiliar real images for synthetic artifacts. We propose EviCal to address both problems by learning complementary detection cues across generators and calibrating their scores using real images as the reference. EviCal uses multiple lightweight experts and measures each expert’s usefulness by how much its contribution reduces the combined detector’s classification loss. This measure guides two training objectives: encouraging each expert to contribute on more than one source generator, and encouraging different experts to improve predictions on different images rather than duplicate one another. After training, EviCal calibrates global and local detection scores using held-out real images from the training sources. It learns how local scores normally relate to global responses on real images, then discounts local scores that are high but consistent with this relationship. Unexpected local deviations can therefore contribute to detection without treating every strong local response as suspicious. Trained on four GenImage generators, EviCal achieves 91.46% accuracy on Chameleon and 95.74% average accuracy on AIGI-Now Pixel, exceeding the strongest evaluated baselines by 12.77% and 3.47%, respectively, while retaining 98.66% average accuracy on GenImage. Cumulative ablations show additional cross-dataset gains from evidence regularization and real-image calibration. The resulting detector operates on individual images without generator identities or test-time adaptation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
ai-generated image detection, multi-evidence learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_k7BGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies cross-generator generalization for AI-generated image detection. Two practical pain points are addressed: 1) multiple LoRA experts tend to learn redundant shortcut cues instead of complementary forensic evidence; 2) high local response scores do not always imply forgery, since real images can also produce strong local activations. The authors propose a two-stage pipeline: multi-source evidence-aware training with generator-consensus loss and evidence-diversity loss; followed by offline real-image-only response calibration after freezing network weights.

Critical strengths
  1. It identifies two under-discussed failure modes in multi-expert detection models: expert redundancy and false alarms from normal real-image local responses.
  2. The proposed evidence diversity regularizer measures expert complementarity via leave-one-out loss increment, instead of naive feature space orthogonality. The real-image-only calibration stage is a practical attempt to suppress false positives.
  3. Strong cross dataset results are reported on challenging benchmarks. Cumulative ablation studies validate the contribution of each main component.
Critical weaknesses
  1. Evidence learning builds entirely upon underlying DINOv3 features. The evidence oriented losses cannot repair bad shortcut representations learned by the backbone. If base features capture spurious correlations, even well-regularized experts cannot produce valid forensic evidence. This risk lacks dedicated ablation.
  2. Calibration is purely post-processing. It only rescales output scores and cannot fix erroneous representations inside network layers. Calibration performance heavily relies on distribution compatibility between held-out real images and test data. Degradation under real image distribution shift is not studied.
  3. The pipeline is relatively heavy with many hyper parameters (expert number, quantile q, etc.). The evidence diversity loss involves determinant computation, which raises potential numerical stability concerns during training.
  4. Contributions are mostly loss design and post processing tricks built on existing LoRA expert framework.
Questions for authors
  1. Evidence learning builds entirely upon underlying DINOv3 features. The evidence oriented losses cannot repair bad shortcut representations learned by the backbone. If base features capture spurious correlations, even well-regularized experts cannot produce valid forensic evidence. This risk lacks dedicated ablation.
  2. Calibration is purely post-processing. It only rescales output scores and cannot fix erroneous representations inside network layers. Calibration performance heavily relies on distribution compatibility between held-out real images and test data. Degradation under real image distribution shift is not studied.
  3. The pipeline is relatively heavy with many hyper parameters (expert number, quantile q, etc.). The evidence diversity loss involves determinant computation, which raises potential numerical stability concerns during training.
  4. Contributions are mostly loss design and post processing tricks built on existing LoRA expert framework.
4 comments

Q1: If base features capture spurious correlations, even well-regularized experts cannot produce valid forensic evidence. This risk lacks dedicated ablation.

A: Thank you for raising this concern. It is important to distinguish between the information encoded by a backbone and the information predominantly used for prediction. A model can rely heavily on spurious correlations while its representations still retain useful alternative cues. This distinction has been demonstrated by Kirichenko et al. [1], who showed that retraining the final classifier can improve shortcut robustness by changing how existing features are used. Therefore, the presence of spurious correlations does not by itself imply that downstream evidence learning cannot improve generalization.

EviCal addresses this feature-utilization problem by learning how shared token information is transformed into evidence responses. Although the base DINOv3 weights are frozen, heterogeneous LoRA updates modify operators inside the Transformer, allowing experts to learn different feature transformations from their shared inputs. The evidence objectives directly optimize these updates, shaping the resulting representations rather than merely regularizing classifiers attached to identical, fixed final embeddings. Generator consensus evaluates an expert’s marginal contribution separately across source generators, favoring usefulness that extends beyond an individual source. Evidence diversity then discourages repeated sample-wise contribution patterns among experts. Together, these objectives are designed to identify complementary discriminative directions supported by multiple sources.

Our claim is therefore about improving the use of potentially transferable information, not recovering information absent from the representation or guaranteeing a universally shortcut-free backbone. We examined the capacity for such improvement in a matched same-backbone control: compared to the original baseline, Evical improves the performance from 82.21% to 86.26% in Chameleon and from 87.63% to 90.26% in AIGI-Now. These results demonstrate that the backbone does leave room for improved transfer through adaptation.

[1] Last layer re-training is sufficient for robustness to spurious correlations. ICLR 2023.

Q2: Calibration limitations.

Thank you for your constructive comment. Our calibration operates after evidence extraction, but it does more than uniformly rescale the outputs. It models how local responses relate to global evidence on source-real images and removes the predictable component before aggregation. The same local score can therefore receive different interpretations depending on the image's global response, helping distinguish unexpected local evidence from high responses that also occur naturally in real images.

We agree that calibration cannot recover information absent from the learned representation. Representational adaptation is handled by the trainable evidence experts, whereas calibration addresses how their responses should be interpreted against real-image behavior. These two components therefore address different aspects of detection.

Regarding distribution compatibility, both the calibration statistics and decision threshold are fitted exclusively on held-out source-real images and remain fixed across all evaluation datasets, without target-based refitting or dataset-specific thresholds. The real-image collections also differ: GenImage uses ImageNet images, Chameleon collects photographs from platforms such as Unsplash, and AIGI-Now Pixel uses COCO test images. Cross-dataset evaluation therefore changes the real-image source as well as the fake generators.

For better show the detailed performance under cross-dataset setting, we test the AUROC and FPR under target Chameleon and AIGI-Now Pixel with source GenImage. The fixed source-fitted calibration retains high real-image accuracy alongside effective fake-image detection on both cross-dataset benchmarks. These results support Evical can keep a good ability transfer across the evaluated real-image collections.

Table 1. Cross-dataset performance with fixed source-fitted calibration. The source FPR target is set to 3%.

DatasetAUROC (%)Source FPR (%)Target-real FPR (%)Fake ACC (%)
Chameleon96.533%3.9385.33
AIGI-Now Pixel99.273%5.0396.51

Q3 (a): The pipeline is relatively heavy with many hyper parameters.

A: Thank you for raising this concern. EviCal does introduce additional hyperparameters, including the number of expert and calibration settings you mentioned. However, the quantities appearing in the calibration procedure are not all independently tuned hyperparameters: the response medians and spreads, regression coefficients, residual-tail offsets, and final decision threshold are estimated from held-out source-real images. Once fitted, the same calibration model and threshold are used across all test datasets, without dataset-specific refitting. The remaining settings are shared across datasets and explicitly specified in the implementation details. This does not eliminate the additional configuration burden, but distinguishes the settings that require specification from the quantities estimated automatically.

Meanwhile, Table 3 of the paper reports sensitivity to the number of experts (K) and calibration quantile (q). Four experts achieve the highest mean accuracy across the three benchmarks. For (q=0.95)–(0.98), mean accuracy varies by only 0.35%, while (q=0.99) reduces cross-dataset accuracy. We also distinguish configuration selection from calibration fitting. The main configuration, (K=4,q=0.97), was selected using mean accuracy across the three benchmarks, so the target-benchmark results informed this choice. We additionally report a configuration selected using GenImage alone, (K=6,q=0.99), which achieves 88.28% on Chameleon and 94.03% on AIGI-Now Pixel. For either configuration, calibration fitting uses only held-out source-real images, without dataset-specific refitting.

Q3 (b): Potential numerical stability concerns during training.

A: Thanks for you attention on this concern. Our implementation evaluates the combined consensus–diversity objective through an algebraically equivalent reuse-weighted Gram formulation, consistent with Eq. (7). We add diagonal regularization of 1e-3 to this weighted matrix to regularize rank-deficient contribution profiles. We compute its log-determinant using the standard PyTorch function torch.linalg.slogdet and use a denominator floor of 1e-4 during profile normalization. Furthermore, we did not observe any non-finite objectives or gradients in our additional loss-ablation experiments, providing further empirical support for the stability of the implementation under the evaluated configuration. These observations support the numerical stability of the implementation under the evaluated settings.

Q4: Contributions are mostly loss design and post processing tricks built on existing LoRA expert framework.

A: Thank you for raising this important question about the methodological contributions. Our contribution is not LoRA itself, but how complementary evidence is defined and learned. EviCal measures each expert’s utility by its marginal reduction in detector loss, then uses this same utility to encourage contributions across generators and on different images. Source-wise centering prevents differences in average generator preference alone from being treated as evidence diversity. This targets useful predictive contributions rather than merely different parameters or features. Calibration also goes beyond score rescaling or threshold adjustment. It models the local response expected from the global and expert state on real images, then removes this predictable component before aggregation. The same local score can therefore be interpreted differently depending on its context, potentially changing image rankings. The methodological contribution lies in contribution-guided evidence learning and conditional interpretation of evidence, with LoRA serving as the implementation framework.

Finally, we sincerely appreciate your positive and thoughtful review which encourages us a lot. We hope our responses have adequately addressed your concerns. If anything remains unclear or requires further clarification, we are willing to provide additional explanations and welcome any further feedback.

Dear Reviewer k7BG,

Thank you for your thoughtful comments. Since the discussion deadline is approaching, we just wanted to check whether there are any remaining issues that would benefit from further clarification. If so, we would be happy to provide additional experiments or explanations before the discussion concludes.

We truly appreciate your time and consideration.

Best regards,

Authors

Dear Reviewer k7BG,

Thank you for reviewing our paper and sharing your suggestions. Your feedback has helped us improve the paper.

As there is little time left in the rebuttal period, we just wanted to check whether our responses have addressed your concerns. If anything remains unclear or unresolved, please let us know, and we will do our best to address it in the remaining time.

Thank you for your time and help!

Best regards,

Authors

Reviewer_J5kvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EviCal, a novel framework for AI-generated image detection. It targets the multi-source setting, where the fake images of the training set are from multiple generators. EviCal utilizes a frozen DINOv3 backbone, and its central components are: (1) Multi-Source Evidence Learning: training heterogeneous experts based on global and local DINOv3 features that provide diverse evidence while remaining useful across the source generators; (2) Real-Image Response Calibration: using the real images to calibrate the detection scores. Experiments suggest the effectiveness of the proposed methods.

Critical strengths
  1. The performance of the proposed method is tested on unseen generators, including images collected in the wild, and the generalization accuracy is impressive.
  2. The study targets multi-source AIGI detection, a practical but less explored scenario, and the proposed Generator-Consensus Loss utilizes the advantage of multiple source generators.
Critical weaknesses
  1. Clarification of the compared methods and the fairness of comparison. Sec. 4.3 states that "All methods are trained on Midjourney, SD1.4, ADM, and BigGAN from GenImage" (line 288). However, many compared methods in Table 1 are not proposed for the multi-source scenario and are not originally trained on all four generators as the proposed method does. In addition, the paper does not clarify the network backbone or pretrained weights these compared methods use. If the results are produced by this paper using aligned training data and the DINOv3 backbone, then the details should be clarified, including how each method is adapted (e.g., how DRCT applies reconstruction on BigGAN images); otherwise, if the results are taken from the original papers, then the comparison is unfair, and the superiority of the proposed method can be attributed to the training data covering multiple generators and the DINOv3 backbone.
Questions for authors
  1. Fig. 1(b) seems to suggest that (1) different experts are encouraged to provide different evidence that is potentially required by different source generators, and (2) real images form one compact cluster instead of many isolated clusters like the fake images. Does point (1) contradict the Generator-Consensus Loss? Is (2) the implicit assumption of the Real-Image Response Calibration?
  2. It seems difficult to tell the differences between the learned experts from the qualitative examples in Fig. 3. Are there any statistical results that can indicate how the experts differ?
4 comments

W1: The fairness of comparison with current method.

A: Thank you for highlighting this ambiguity. All compared baselines were retrained by us using their publicly available official implementations and the same four-source image pool as EviCal: Midjourney, SD1.4, ADM, and BigGAN from GenImage. All reported results are obtained through our own reproduction and evaluation, rather than directly taken from the original papers. Thus, access to multiple source generators is shared across the compared methods, rather than being an advantage exclusive to EviCal.

We retain each baseline’s method-specific backbone and pretrained initialization. For example, DRCT uses ConvNeXt-Base, UniversalFakeDetect uses CLIP ViT-L/14, and PGC uses DINOv2-Large. The main comparison uses the decision thresholds or decision rules specified by the respective official implementations. We will clarify this protocol in Sec. 4.3.

Regarding the reproduction of DRCT mentioned in your review, we use the fixed stable-diffusion reconstructor to process source images, including BigGAN outputs, as ordinary RGB inputs. The detector is trained with four image groups: original real images, reconstructed real images, original fake images, and reconstructed fake images. Original real images receive the real label; the other three groups receive the fake label, following DRCT’s training formulation.

Q1: 1. Fig. 1(b) seems to suggest that (1) different experts are encouraged to provide different evidence that is potentially required by different source generators, and (2) real images form one compact cluster instead of many isolated clusters like the fake images. Does point (1) contradict the Generator-Consensus Loss? Is (2) the implicit assumption of the Real-Image Response Calibration?

A: Thank you for raising these questions. For point (1), different experts are intended to provide complementary evidence, without a one-to-one correspondence between experts and generators. The Generator-Consensus Loss encourages each expert’s useful contributions to extend beyond a single source generator, while the Evidence-Diversity Loss discourages experts from duplicating the same contribution patterns across images. These objectives are compatible: different experts can capture different cues, with each cue being useful across multiple generators. Their contribution strengths may still vary across generators and samples. Consensus therefore concerns cross-generator usefulness, rather than identical responses or equal contributions.

For point (2), Real-Image Response Calibration does not assume that real images form a single compact cluster, but relies on source-real response statistics and relationships that remaining useful for unseen real images. Our motivation is that, compared with different fake distributions arising from diverse generative mechanisms, real-image responses may provide a relatively stable calibration reference. Specifically, RC estimates from source-real images. Its cross-domain usefulness depends on these quantities remaining informative for target-real images. Real images can still exhibit multiple modes associated with content, cameras, and processing. Neither the calibration formulation nor the training objective requires them to collapse into one cluster. Thus, the compact real region in Fig. 1(b) represents a real-referenced decision region, rather than a literal unimodal feature distribution. Relative stability is a potential working hypothesis motivating this reference, not a guarantee that real-image distributions or false-positive rates remain unchanged across datasets.

Considering the potential misleading , following your suggestion, we have revised the Fig. 1 to clarify that the idea of Evical. It implies neither a one-to-one expert–generator correspondence nor a single compact real-image cluster.

Q2: Are there any statistical results that can indicate how the experts differ?

A: Thank you for this helpful suggestion. Following your suggestion, we conducted an additional leave-one-expert-out analysis to examine the contributions of individual experts. We also add this experiments in the main paper. As shown below, removing any one of the four experts reduces accuracy on both Chameleon and AIGI-Now Pixel. The full model achieves the highest accuracy on both cross-dataset benchmarks, supporting the complementary contributions of all four experts to cross-dataset detection. Their effects also differ across datasets: removing E4 causes the largest degradation on all three benchmarks, while removing E3 affects Chameleon more than removing E2, with the reverse holding on AIGI-Now Pixel. Although removing E1 and E3 improves GenImage accuracy by 0.46% and 0.40%, respectively, both removals consistently reduce performance on the two cross-dataset benchmarks. It may indicate that the remaining experts may already capture the dominant discriminative cues, while additional responses from E1 and E3 can occasionally shift borderline predictions in an unfavorable direction, while for cross-dataset, those strong evidence in GenImage may less reliable, making complementary responses from different experts more valuable for samples where other branches provide weak evidence.. Their inclusion therefore involves a small trade-off in GenImage accuracy while benefiting cross-dataset detection. This is consistent with our objectives, which encourage useful and nonredundant contributions across training generators. The results demonstrate both the different contributions of individual experts and the benefit of combining all four for cross-dataset detection.

DatasetFull Modelw/o E1w/o E2w/o E3w/o E4
GenImage98.6699.1296.8799.0694.82
Chameleon91.4689.6890.8488.3487.97
AIGI-Now Pixel95.7495.0093.0194.0390.11

Finally, we sincerely appreciate your positive and thoughtful review which encourages us a lot. We hope our responses have adequately addressed your concerns. If anything remains unclear or requires further clarification, we are willing to provide additional explanations and welcome any further feedback.

Thank you for the valuable response, which has addressed most of my concerns. Here are some further comments:

  • The clarified retraining protocol makes sense, but the comparison fairness regarding the backbone is still concerning. From my point of view, this is not a critical weakness sufficient to support rejection of the paper, especially given the effectiveness of the proposed components as verified by the updated ablation studies. Nevertheless, the paper should clarify this to avoid misinterpretation of the results.
  • The leave-one-expert-out ablation provides some additional information about how the experts differ, although dataset-wise statistics may not give us intuitive explanations, like which expert captures which kind of clues. Again, this is not a major weakness of the paper, while further interpretations may be expected if feasible. Besides, I recommend a notation of the accuracy drop in this table to make the differences easier to read.

Overall, I am leaning towards acceptance, and will keep my rating of 3.

Thank you for your support and for taking the time to follow up on our responses. Your feedback has been very encouraging to us. We have revised Sec. 4.3 to explain our retraining protocol and have also added the accuracy drops to the ablation table, as you suggested. Due to the limited time remaining, we may not be able to finish the further experiments within the rebuttal period, but we will continue working on it and include any additional findings in the camera-ready version.

Best wishes!

Reviewer_ZdYbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EviCal for AI-generated image detection under generator and dataset shifts. The method combines heterogeneous LoRA experts on a frozen DINOv3 backbone with shallow local responses. Expert learning is regularized using the change in classification loss when an expert’s contribution is removed, encouraging usefulness across multiple generators and complementary contributions across images. A subsequent calibration stage uses held-out source-real images to normalize responses, predict local scores from the global state, and set a fixed decision threshold. Trained on four GenImage generators, EviCal reports 98.66% average accuracy on GenImage, 91.46% on Chameleon, and 95.74% on AIGI-Now Pixel.

Critical strengths
  1. The cross-dataset results are strong. In Table 1, EviCal exceeds the best evaluated baseline by 12.77 percentage points on Chameleon and 3.47 points on AIGI-Now Pixel, while retaining high GenImage accuracy. Evaluation on both benchmarks provides useful evidence of transfer beyond the training generators.
  2. The method gives a concrete definition of expert usefulness through its contribution to classification loss. Using this quantity for both cross-generator reuse and image-level diversity is well motivated. The conditional calibration also addresses a relevant issue: a large local response need not indicate manipulation if it is predictable from the detector’s behavior on real images.
Critical weaknesses
  1. The ablations do not isolate the main methodological contributions. Table 2 adds the consensus and diversity losses together, then introduces the entire calibration stage. It therefore does not establish the separate effects of the two losses or the benefit of conditional residual correction over simpler real-only normalization and threshold calibration. This matters because calibration alone adds 5.20 and 5.48 percentage points on the two cross-dataset benchmarks. Table 3(c) compares response pathways, but does not resolve this attribution issue.
  2. The evaluation does not adequately separate discrimination quality from operating-point selection. Only accuracy is reported, and the threshold-selection procedures for the competing methods are not specified. Since EviCal explicitly fits a threshold using held-out real images, the comparison would be more convincing with AUROC, separate real/fake accuracies, and results at a common source-real false-positive rate. Applying comparable source-real threshold calibration to representative strong baselines would help establish whether the reported advantage persists beyond threshold adjustment.
Questions for authors
  1. With the evidence network fixed, how do normalization plus source-real threshold calibration, unconditional local tail correction, and the full conditional residual correction compare? This would clarify how much of the calibration gain comes from modeling global–local dependencies.
  2. What threshold-selection procedure was used for each baseline? Does EviCal retain its advantage in AUROC and when representative baselines are calibrated to the same source-real false-positive rate?
  3. How were the main configuration and calibration hyperparameters selected, including q, the ridge coefficient, and the number of experts? Please identify the validation data used and confirm whether Chameleon or AIGI-Now results influenced these choices.
5 comments

W1&Q1: The ablations about the consensus and diversity losses

A: Thank you for raising this important concern. Following your suggestion, we have conducted two additional controlled ablations.

First, we evaluate GCL and EDL independently. Starting from the same baseline (96.34/82.21/87.63), GCL achieves 97.71/85.89/89.18, while EDL achieves 98.93/84.61/88.67 on GenImage, Chameleon, and AIGI-Now, respectively. Combining both losses further improves cross-dataset performance to 86.26% on Chameleon and 90.26% on AIGI-Now, suggesting their complementary contributions.

Second, we isolate the calibration components with the evidence network fixed. Starting from raw scores with source-real threshold calibration, normalization improves Chameleon/AIGI-Now accuracy to 87.96%/93.58%. Unconditional local tail correction further increases them to 89.89%/95.16%, while full conditional residual correction achieves 91.46%/95.74%. Notably, conditional correction provides additional gains over unconditional tail correction, supporting the benefit of explicitly modeling global–local dependencies beyond simple score normalization and tail adjustment.

These additional experiments provide a clearer attribution of the individual contributions and validate the effectiveness of the proposed evidence learning and conditional calibration.

Table 1: Ablation of the GCL and EDL.

MethodGenImageChameleonAIGI-Now
Base96.3482.2187.63
+Generator-Consensus Loss (GCL)97.7185.8989.18
+Evidence-Diversity Loss (EDL)98.9384.6188.67
+EDL+ GCL98.8886.2690.26

Table 2: Ablation of the Real-Image Response Calibration.

MethodGenImageChameleonAIGI-Now
Raw + Threshold98.8886.2690.26
+ Normalization98.7287.9693.58
+ Unconditional local tail correction98.7489.8995.16
Full Real-Image Response Calibration98.6691.4695.74

W2&Q2: Performance under same source-real false-positive rate.

A: Thank you for your constructive suggestions. All baselines are trained using their publicly available official implementations on the same training set as EviCal. In the original evaluation, each baseline uses the default decision threshold specified in its official implementation.

Following your suggestions, we additionally report AUROC, overall accuracy, and separate real/fake accuracies, and evaluate all five baselines under a common source-real calibration protocol (We retrained TranX-Adapter after the original checkpoint was deleted. Results in Tables 3–5 therefore differ slightly from those in Table 1 of the manuscript.). Specifically, we calibrate each method’s threshold on the same source-real calibration set used by EviCal, targeting 3% FPR, where a false positive denotes a real image classified as fake. Each method’s calibrated threshold is then fixed across all test datasets, without target-specific threshold tuning. The results under the original thresholds and the common source-real FPR are presented in Tables 3–5. Since FPR is matched on the source calibration set, real-image accuracies can differ on the test datasets. Overall, EviCal maintains its cross-dataset advantage in both AUROC and accuracy under the common calibration protocol. These results demonstrate that its generalization capability is not merely attributable to threshold selection or operating-point adjustment, but reflects stronger discrimination between real and AI-generated images across unseen datasets.

Table 3: Performance of EviCal on GenImage (In-dataset)

MethodAUROCOrigianl (ACC / Real / Fake)Same source-real FPR (ACC / Real / Fake)
IAPL(CVPR'26)99.9999.76 / 99.73 / 99.7998.54 / 97.11 / 99.99
LTD(CVPR'26)99.6397.29 / 97.33 / 97.2797.24 / 96.88 / 97.63
PGC(ICML'26)100.0099.72 / 99.53 / 99.9297.82 / 95.71 / 99.99
OmniAID (ICML'26)96.3589.03 / 84.57 / 93.4387.91 / 96.63 / 78.69
TranX-Adapter (ICML'26)98.2089.38 / 99.14 / 79.5492.59 / 96.25 / 88.95
EviCal99.9998.66 / 97.38 / 99.9898.66 / 97.38 / 99.98

Table 4: Performance of EviCal on Chameleon (Cross-dataset)

MethodAUROCOrigianl (ACC / Real / Fake)Same source-real FPR (ACC / Real / Fake)
IAPL(CVPR'26)70.1960.00 / 99.82 / 6.9964.05 / 89.69 / 29.92
LTD(CVPR'26)76.1663.17 / 98.11 / 16.6663.67 / 97.79 / 18.26
PGC(ICML'26)62.6362.64 / 99.10 / 14.1065.57 / 88.97 / 34.41
OmniAID (ICML'26)83.1769.87 / 53.06 / 92.2475.62 / 80.93 / 68.54
TranX-Adapter (ICML'26)89.6277.97 / 68.06 / 91.1673.05 / 57.17 / 94.18
EviCal96.5391.46 / 96.07 / 85.3391.46 / 96.07 / 85.33

Table 5: Performance of EviCal on AIGI-Now Pixel (Cross-dataset)

MethodAUROCOrigianl (ACC / Real / Fake)Same source-real FPR (ACC / Real / Fake)
IAPL(CVPR'26)86.1777.61 / 99.99 / 55.2279.08 / 97.16 / 61.00
LTD(CVPR'26)93.9387.13 / 96.79 / 77.4887.34 / 96.22 / 78.47
PGC(ICML'26)88.1877.25 / 99.40 / 55.1085.18 / 97.23 / 73.13
OmniAID (ICML'26)87.0179.25 / 76.99 / 81.5172.34 / 94.67 / 50.02
TranX-Adapter (ICML'26)98.7191.53 / 98.99 / 84.0794.30 / 96.08 / 92.52
EviCal99.2795.74 / 94.97 / 96.5195.74 / 94.97 / 96.51

Q3: How were the main configuration and calibration hyperparameters selected?

A: Thank you for requesting this clarification. We have revised Table 1 of the paper to explicitly report two configuration-selection criteria. EviCal° ((q=0.99, K=6)) selects these hyperparameters using accuracy on the GenImage, without using Chameleon or AIGI-Now results. EviCal† ((q=0.97, K=4)) uses the configuration selected by mean accuracy across the three benchmarks. Chameleon and AIGI-Now results therefore informed the latter selection, which we now distinguish explicitly from the GenImage-only setting.

Calibration fitting uses 1000 held-out source-real images, divided into two disjoint sets: 500 for response normalization, conditional prediction, and residual-tail estimation, and 500 for determining the final decision threshold. For each fixed configuration, these calibration quantities and the final decision threshold are estimated exclusively from the held-out source-real images. Chameleon and AIGI-Now images are not included in either calibration subset. However, their evaluation accuracies inform the selection of q and K for EviCal† in Table 1 of the paper. We report results for both settings in Table 1 of the paper: EviCal° (q=0.99, K=6), selected using GenImage accuracy alone, and EviCal† (q=0.97, K=4), selected using mean accuracy across the three benchmarks. The ridge coefficient is (\lambda=0.1), determined by the accuracy of GenImage without Chameleon or AIGI-Now. We have added this to the implemental details. For each configuration, the fitted calibration quantities and decision threshold are fixed across all evaluation datasets.

The additional GenImage-selected configuration achieves 88.28% accuracy on Chameleon and 94.03% on AIGI-Now, exceeding the strongest respective baselines reported in Table 1 of the paper. These results support our main finding that EviCal improves cross-dataset detection, with the advantage remaining when q and K are selected using GenImage alone.

Finally, we sincerely appreciate your careful review and constructive feedback. We hope our responses have adequately addressed your concerns. If anything remains unclear or requires further clarification, we are willing to provide additional explanations and welcome any further feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.