acceptodds
Under review as a conference paper at ICLR 2027

Relative Patch Response Learning for Generalizable AI-Generated Image Detection

Abstract

Generative models can now synthesize highly realistic images, simultaneously increasing the risks of misinformation and visual forgery. Therefore, detecting AI-generated images becomes more essential, and a reliable detector must generalize to unseen generators and stay robust to unseen perturbations in the wild. Existing detectors are typically trained on either independently collected real and generated images or aligned real-generated pairs designed to mitigate content bias. Building on aligned pairs, recent methods form a mixed view by replacing some patches of the real image with their generated counterparts. However, we find that self-attention lets real and generated patches interact, so the feature of each patch no longer reflects its own source alone. This contextual shift makes a per-patch source label an imprecise target. To this end, we propose Relative Patch Response Learning (PRL). Instead of labeling each patch, PRL compares the same patch across two mixed views of an aligned pair and learns from its patch response, the change of its score between the views. (i) To give precise supervision under the contextual shift, a relative response objective measures the responses of source-changed patches against those of source-unchanged patches, which respond to the shift alone. (ii) To provide a reliable reference for the shift, a reference coherence objective keeps each group of source-unchanged patches moving as a whole. (iii) Since the two views contain different amounts of generated content, an area ranking objective asks the view with the larger generated area to have a higher mean patch score. Extensive experiments demonstrate the superior performance of PRL, which surpasses the best prior methods by 4.3% and 5.9% in average balanced accuracy across eight standard and three in-the-wild benchmarks, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, AI safety

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_GNacOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Relative Patch Response Learning (PRL) to improve the generalization of AI-generated image detectors. It identifies contextual shift in vision transformers: the representation or score of a patch can change when surrounding regions change, even if the patch itself remains unchanged. To address this issue, PRL constructs two mixed views from a real image and its aligned synthetic counterpart, compares patch scores at corresponding positions, and uses source-unchanged patches as references to estimate and reduce contextual effects. Training combines relative response, reference coherence, and area ranking objectives with image-level classification. The paper reports strong overall performance on cross-generator, in-the-wild, and local manipulation detection benchmarks. Its main contribution is a patch-level supervision strategy that accounts for contextual effects to better learn traces of image generation.

Critical strengths

1.The supervision design accounts for contextual effects. The paper highlights that self-attention can cause unchanged patches to respond to changes elsewhere in the image and illustrates this phenomenon through visualizations. This provides a clear motivation for revisiting per-patch source supervision on mixed views. 2.The method makes better use of correspondence within aligned image pairs. It compares scores at corresponding patch positions across two mixed views and uses source-unchanged patches as references. Compared with independently classifying aligned images, this design exploits finer-grained spatial correspondence and aims to reduce interference from content differences and contextual changes. 3.The evaluation is broad and the overall results are strong. Experiments cover cross-generator, in-the-wild, and local manipulation detection, supporting the method’s generalization ability and potential applicability across different detection scenarios. 4.The ablation studies provide useful controlled comparisons. The paper compares different supervision strategies under the same backbone and training data, examines the contribution of each auxiliary objective, and includes a same-backbone comparison with DDA. These experiments help distinguish the benefits of the supervision design from those of the backbone and support the contributions of the individual components.

Critical weaknesses

The gains from the proposed supervision are not fully isolated from those of the backbone and training configuration. The main table evaluates official checkpoints using their respective inference pipelines, with differences in backbones, training data, and preprocessing configurations. While this comparison is informative about the performance of complete detection systems, the full performance margin cannot be attributed solely to PRL’s supervision mechanism. The paper does provide useful controls: in Table 13, adapting DDA from its original DINOv2 configuration to DINOv3 with LoRA improves its average result from 89.16% to 93.20%, compared with 95.67% for PRL, indicating a substantial contribution from the backbone and adaptation configuration.

Questions for authors

1.All PRL experiments use DINOv3 ViT-L/16. Does the proposed supervision also benefit vision encoders trained with different pretraining objectives, such as the CLIP vision encoder? A comparison between PRL and an image-level BCE baseline under otherwise identical settings on another ViT backbone would help establish whether the gains generalize beyond DINOv3. 2.Please clarify the decision threshold used to compute balanced accuracy and how it was selected. Was a fixed threshold used across all test benchmarks? If threshold selection or calibration was performed, which validation data were used, and did the procedure involve any information from the target test sets? Reporting AUROC and average precision (AP) for the synthetic class would help distinguish ranking performance from the effects of threshold selection.

3 comments
AuthorsOpenReview ↗

Author Response (Part 1/2)

We thank the reviewer for the detailed review and the constructive comments. We also thank the reviewer for recognizing the motivation and visualization of the contextual shift, our use of the position-wise correspondence within aligned pairs, the broad evaluation across cross-generator, in-the-wild and local manipulation settings, and the controlled ablations.

W1: The gains from the proposed supervision are not fully isolated from those of the backbone and training configuration. The main table evaluates official checkpoints using their respective inference pipelines, with differences in backbones, training data, and preprocessing configurations. While this comparison is informative about the performance of complete detection systems, the full performance margin cannot be attributed solely to PRL’s supervision mechanism. The paper does provide useful controls: in Table 13, adapting DDA from its original DINOv2 configuration to DINOv3 with LoRA improves its average result from 89.16% to 93.20%, compared with 95.67% for PRL, indicating a substantial contribution from the backbone and adaptation configuration.

Q1: All PRL experiments use DINOv3 ViT-L/16. Does the proposed supervision also benefit vision encoders trained with different pretraining objectives, such as the CLIP vision encoder? A comparison between PRL and an image-level BCE baseline under otherwise identical settings on another ViT backbone would help establish whether the gains generalize beyond DINOv3.

We agree that the backbone contributes to the overall performance. We further evaluate our PRL on CLIP ViT-L/14, which uses a different pretraining objective, to demonstrate the effectiveness of our proposed method beyond DINOv3. Together with the existing controlled comparisons, these results provide three pieces of evidence:

  1. Same backbone, same training data. Using the same DINOv3 backbone and LoRA adaptation settings, the image-level BCE baseline reaches 92.65% and PRL reaches 95.55% (+2.90%). Adding mixed views under the image-level label gives 92.68%, and a per-patch source label on the mixed views gives 91.82% (-0.83%). These results demonstrate that the improvement comes from how the mixed views are supervised.
  2. Comparison with DDA on the same backbone. As the reviewer notes, transferring DDA from DINOv2 to DINOv3 with LoRA improves its average from 89.16% to 93.20% (+4.04%). This gain reflects the contribution of the backbone and adaptation configuration. With the same backbone and aligned training pairs, PRL reaches 95.67%, outperforming DDA by 2.47% across the 11 benchmarks.
  3. Generalization to CLIP. Following Q1, we additionally evaluate PRL and the image-level BCE baseline on CLIP ViT-L/14 under otherwise identical settings. PRL improves average BAcc from 72.97% to 84.39% (+11.42%) across the 11 benchmarks, with gains of 12.27% on the eight standard benchmarks and 9.17% on the three in-the-wild benchmarks. (Diff. = PRL − image-level BCE)
CLIP ViT-L/14GenImageAIGCDetectDRCT-2MDDA-COCOEvalGENSynthbusterForenSynthsUnivFDChameleonSynthWildXWildRFStandardWildAvg.
Image-level BCE89.8983.4399.5056.8182.1563.0768.2366.7961.8568.4562.5376.2364.2872.97
PRL96.5191.8595.9979.1698.9388.0182.8474.6870.4873.1076.7788.5073.4584.39
Diff.+6.62+8.42−3.51+22.35+16.78+24.94+14.61+7.89+8.63+4.65+14.24+12.27+9.17+11.42

Further comparison with DDA on CLIP. The DDA paper reports CLIP ViT-L/14 results on only five benchmarks, and the corresponding checkpoint and training code have not been released. We therefore evaluate PRL using CLIP ViT-L/14 on these same five benchmarks and compare against DDA’s reported results. PRL outperforms DDA by 3.78% on average. (Diff. = PRL - DDA)

MethodGenImageDRCT-2MEvalGENChameleonSynthWildXAvg.
DDA (CLIP ViT-L/14)97.0080.4099.2067.7071.8083.22
PRL (CLIP ViT-L/14)96.5195.9998.9370.4873.1087.00
Diff.−0.49+15.59−0.27+2.78+1.30+3.78

We add the results on the CLIP backbone to the appendix of the revised paper (Appx. E; Table 23).

AuthorsOpenReview ↗

Author Response (Part 2/2)

Q2: Please clarify the decision threshold used to compute balanced accuracy and how it was selected. Was a fixed threshold used across all test benchmarks? If threshold selection or calibration was performed, which validation data were used, and did the procedure involve any information from the target test sets? Reporting AUROC and average precision (AP) for the synthetic class would help distinguish ranking performance from the effects of threshold selection.

We thank the reviewer for raising this point, which helps evaluate the detector more completely.

Threshold. Following the standard practice in AI-generated image detection, our detector outputs a probability score y^∈[0,1]\hat{y} \in [0, 1] indicating the likelihood that an image is synthetic. We use a fixed threshold of 0.5, classifying an image as synthetic if y^≥0.5\hat{y} \geq 0.5. This same threshold is applied uniformly across all evaluated benchmarks to compute balanced accuracy. We have clarified the decision threshold in Section 4.1 of the revised paper.

AUROC and AP. The tables below report AUROC and AP, with synthetic images as the positive class, for PRL and all compared methods on the benchmarks of the paper. EvalGEN is excluded because it contains only synthetic images. PRL achieves the best AUROC and AP on all 10 benchmarks, with averages of 98.83 and 98.55. This further supports the effectiveness of PRL.

AUROC (%):

MethodGenImageAIGCDetectDRCT-2MDDA-COCOSynthbusterForenSynthsUnivFDChameleonSynthWildXWildRFAvg.
UnivFD83.8187.9583.4868.4153.8692.0592.6855.3746.1353.0871.68
FatFormer63.9476.6031.2348.6651.5895.8292.3358.6058.1572.7564.97
NPR38.7234.3845.2248.9038.0323.7726.6732.4937.8460.7838.68
DRCT97.5476.8399.2980.2274.6955.2669.8874.6482.4680.8179.16
C2P-CLIP81.1684.8382.3256.6062.4497.0395.9739.9754.9060.2171.55
AIDE84.7569.7089.0452.9864.7845.0557.9372.5165.7168.3267.08
AlignedForensics97.5180.5899.8797.5183.0664.9874.7884.7485.1086.2285.43
Effort95.3686.5385.7257.0687.4367.5777.5368.7763.7079.1376.88
SAFE53.6946.9445.4050.0059.5939.6141.4557.1349.1362.6350.56
DDA98.7096.7799.8998.7595.6792.2792.7091.9496.3296.1495.92
GAPL99.1899.3799.6095.7296.5796.4297.5177.0293.7494.6494.98
PRL (Ours)99.7199.8099.9799.1999.6897.8698.4795.5198.4899.6398.83

AP (%):

MethodGenImageAIGCDetectDRCT-2MDDA-COCOSynthbusterForenSynthsUnivFDChameleonSynthWildXWildRFAvg.
UnivFD82.5487.1582.6567.5060.2991.7892.6847.6852.1157.3972.18
FatFormer68.0379.7543.4052.2655.8196.3993.6857.8058.0773.8967.91
NPR42.9343.1347.0347.5344.8040.9041.8539.6443.7065.2845.68
DRCT96.6576.7799.4483.7577.1955.5069.4471.0083.7884.6679.82
C2P-CLIP82.2285.8782.2758.1271.3297.5796.6340.1956.0064.2573.44
AIDE86.2769.6690.3855.5867.1148.2760.3665.3665.0272.0468.00
AlignedForensics96.0981.2299.8997.2881.8166.2374.8482.8587.2490.1385.76
Effort93.6582.9080.1355.8187.6765.3676.6458.5761.1781.4074.33
SAFE56.1350.1048.3352.3657.3545.7446.8650.6449.7866.4652.38
DDA98.4896.4599.9098.8295.8591.2792.6191.2096.3996.7695.77
GAPL99.1599.3499.6195.7396.7896.2197.3674.8394.3494.9694.83
PRL (Ours)99.6499.7999.9799.2599.5597.7698.4193.4898.1099.5998.55

In the revised paper, we add the overall AUROC and AP to the main text (Table 2) and give the per-subset results in the appendix (Appx. D).

AuthorsOpenReview ↗

Follow-up on Author Response

Dear Reviewer GNac, thank you again for your constructive review. To address your concern about the contribution of the backbone and training configuration, our rebuttal highlights controlled comparisons using the same DINOv3 backbone, LoRA settings and training data (W1). Following your suggestion, we also added a controlled comparison on CLIP ViT-L/14, where PRL improves average BAcc from 72.97% to 84.39% across 11 benchmarks, supporting its effectiveness beyond DINOv3 (Q1). We further clarified the fixed threshold of 0.5 across benchmarks and reported AUROC and AP to assess ranking performance separately from threshold selection (Q2). We would appreciate it if you could let us know whether our responses have addressed your concerns.

Reviewer_wHneOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores the task of detecting AI-generated images with a focus on generalization to unseen generators. The goal is to develop a reliable detector capable of generalizing to unseen generative models while remaining robust against various unknown perturbations encountered in real-world applications. To this end, this paper proposes "Relative Patch Response Learning" (PRL). The core motivation is that, in Vision Transformer (ViT) architectures, contextual interactions mean that altering certain image patches can influence the predictions for the remaining, unchanged patches. This paper posits that this characteristic serves as a cue for training a robust detector. Specifically, PRL constructs hybrid views from aligned real and synthetic images and compares their predictions at the patch level. By leveraging the unchanged patches as a reference to estimate contextual shifts, the method incorporates a relative response loss, a reference consistency loss, and a regional ranking loss. Experimental results across eleven benchmark datasets demonstrate that this approach outperforms existing detection techniques.

Critical strengths
  • This paper is well-motivated. The paper identifies a meaningful limitation of existing patch-level supervision. Due to contextual interactions in Vision Transformers, the prediction of a patch may change even when its content remains unchanged. This observation is sound to me

  • The relative response formulation is good. Instead of directly assigning real/fake labels to individual patches, PRL compares their responses across two mixed views. Using unchanged patches as references is an intuitive way to account for contextual changes and better capture source-related differences. This is well aligned with the motivation

  • Good experimental performance. PRL achieves the best average performance across eleven benchmarks, with good improvements on challenging in-the-wild datasets such as Chameleon and WildRF. Furthermore, the paper examines the contributions of its training objectives and compares relative response learning with conventional patch-level supervision.

Critical weaknesses
  • The reference coherence loss may impose an overly strong constraint. The method encourages unchanged patches to have similar response changes. However, patches containing different semantic content may naturally respond differently to changes elsewhere in the image. For me, it is unclear why these responses should be similar or whether enforcing this constraint could suppress useful information.

  • Simpler alternatives comparison. The paper compares PRL with direct patch-level supervision, but it would be useful to include simpler correspondence-based objectives. For example, the authors could compare PRL with a simpler variant that directly supervises the score changes of source-changed patches, without subtracting the responses of unchanged patches.

  • Please discuss limited generalization to certain types of manipulation. Although PRL achieves strong overall results, its balanced accuracy on the DeepFake face-swap subset of ForenSynths is only 55.26%, substantially below several competing methods. Please discuss this potential limitation.

Questions for authors
  • Different patches may interact with their surroundings differently, depending on their content and position. It is unclear whether subtracting the average response of unchanged patches can reliably isolate the effect of changing a patch from real to synthetic.

  • It could be better to discuss recent works on improving cross-generator generalization: Where Detectors Fail: Probing Generative Space for Generalizable AI-Generated Image Detection (ICML 2026), PGC: Peak-Guided Calibration for Generalizable AI-Generated Image Detection (ICML 2026), and Exploiting the Source-Asymmetry Confidence Gap for Generalizable AI-Generated Image Detection (CVPR 2026).

4 comments
AuthorsOpenReview ↗

Author Response (Part 1/3)

We thank the reviewer for the detailed review and the expert comments. We also thank the reviewer for recognizing our observation of contextual interaction in ViTs, the design of the relative response, and the results on in-the-wild benchmarks such as Chameleon and WildRF.

W1: The reference coherence loss may impose an overly strong constraint. The method encourages unchanged patches to have similar response changes. However, patches containing different semantic content may naturally respond differently to changes elsewhere in the image. For me, it is unclear why these responses should be similar or whether enforcing this constraint could suppress useful information.

We thank the reviewer for this thoughtful comment. Different patches indeed contain different semantic content. However, AI-generated image detection aims to identify generation traces rather than rely on semantic content. For example, detectors trained with image-level BCE assign the same binary source label to images regardless of their content, encouraging the learning of generation-related evidence that generalizes across different semantics.

Reference coherence introduces a task-driven regularization on detection-related score changes. Its motivation and scope are as follows:

  1. The patch scoring head is trained to capture generation-related evidence. Reference coherence encourages consistency in the changes of these scores within each reference group. It does not require patch features to be identical.
  2. Content alignment provides a controlled reference. For each source-unchanged patch, the local input is identical across the two views, so taking the score difference already reduces the influence of its local content before applying reference coherence. We regularize the real and synthetic reference groups separately to obtain more coherent group-level baselines for relative response learning.
  3. Ablation results support this design. Figure 6 evaluates different weights of the reference coherence loss. Increasing the weight from 0.1 to 0.3 improves average BAcc from 94.53% to 95.55%. All tested nonzero weights (94.47%–95.55%) outperform removing the loss entirely (94.05%). These results support the effectiveness of reference coherence for cross-benchmark generalization.
λcoh\lambda_{\mathrm{coh}}0 (w/o Lcoh\mathcal{L}_{\mathrm{coh}})0.10.20.3 (default)0.40.5
Avg.94.0594.5394.4795.5594.8794.55
AuthorsOpenReview ↗

Author Response (Part 2/3)

W2: Simpler alternatives comparison. The paper compares PRL with direct patch-level supervision, but it would be useful to include simpler correspondence-based objectives. For example, the authors could compare PRL with a simpler variant that directly supervises the score changes of source-changed patches, without subtracting the responses of unchanged patches.

Q1: Different patches may interact with their surroundings differently, depending on their content and position. It is unclear whether subtracting the average response of unchanged patches can reliably isolate the effect of changing a patch from real to synthetic.

We thank the reviewer for the suggestion and question. We address both points through the motivation for reference subtraction and an additional controlled experiment.

1. The role of reference subtraction (Q1). Source-changed patches respond to both their own source changes and contextual changes, whereas source-unchanged patches provide observations of the contextual response. Content- and position-dependent interactions can vary across patches, this subtraction provides a group-level correction for shared contextual effects, without requiring identical contextual responses at every position. It allows the relative response to better reflect changes in generation traces.

2. Additional experiment and results (W2). Following the reviewer’s suggestion, we evaluate a simpler variant that directly supervises the mean response of source-changed patches without reference subtraction. This variant reduces average BAcc from 95.55% to 92.65%, below even removing the relative response loss entirely (93.03%). These results demonstrate that directly supervising uncorrected responses is less effective, while reference subtraction substantially improves cross-benchmark generalization by accounting for shared contextual shifts.

MethodGenImageAIGCDetectChameleonWildRFAvg.
PRL (Full)96.8397.7590.0197.5995.55
Lrel\mathcal{L}_{\mathrm{rel}} without reference subtraction (new)96.5596.7581.6495.6692.65
w/o Lrel\mathcal{L}_{\mathrm{rel}}96.3296.7783.2095.8293.03
w/o Lcoh\mathcal{L}_{\mathrm{coh}}96.0596.4386.8396.8994.05
w/o Lrank\mathcal{L}_{\mathrm{rank}}95.7297.0586.6197.2094.14

We add the result to the appendix of the revised paper (Appx. E; Table 21).

AuthorsOpenReview ↗

Author Response (Part 3/3)

W3: Please discuss limited generalization to certain types of manipulation. Although PRL achieves strong overall results, its balanced accuracy on the DeepFake face-swap subset of ForenSynths is only 55.26%, substantially below several competing methods. Please discuss this potential limitation.

We thank the reviewer for pointing out this limitation. PRL indeed performs relatively poorly on the DeepFake. This subset originates from FaceForensics++, where generated faces are blended into real video frames to replace identities. We attribute this limitation partly to the substantial distribution gap between such face manipulations and our training data.

We further evaluate PRL on this subset and obtain an AUROC of 81.11% and an AP of 80.53%. These results indicate that PRL retains reasonable ranking ability, while the fixed decision threshold contributes to its low BAcc. As an exploratory analysis, adjusting the threshold from 0.5 to 0.32 improves BAcc from 55.26% to 73.05%.

Future work. We will further improve generalization to such manipulations by expanding the training data and exploring post-hoc threshold calibration methods, such as Exploiting the Source-Asymmetry Confidence Gap for Generalizable AI-Generated Image Detection (SACG, CVPR 2026), as suggested by the reviewer in Q2.


 

Q2: It could be better to discuss recent works on improving cross-generator generalization: Where Detectors Fail: Probing Generative Space for Generalizable AI-Generated Image Detection (ICML 2026), PGC: Peak-Guided Calibration for Generalizable AI-Generated Image Detection (ICML 2026), and Exploiting the Source-Asymmetry Confidence Gap for Generalizable AI-Generated Image Detection (CVPR 2026).

We thank the reviewer for the recommendation. These three works improve cross-generator generalization at three different stages: training data, decision aggregation and post-hoc thresholding. They are orthogonal to PRL. We have discussed them in the Related Work section and Appendix of the revised paper.

  • Where Detectors Fail (PROBE) uses feedback from the detector to guide a generative model toward hard samples that the detector misclassifies. It then fine-tunes the detector on these samples to improve generalization to unseen generators.
  • PGC observes that local generation traces are easily hidden when features are aggregated globally. It highlights local regions with high responses to calibrate the global prediction, which improves cross-generator generalization.
  • Source-Asymmetry Confidence Gap (SACG) exploits the gap in prediction confidence between real images and images from unseen generators. It improves cross-generator generalization with a training-free confidence calibration strategy.

Additional quantitative comparison. We further compare PROBE, PGC and PRL under the evaluation protocol used in our paper. SACG is not included because we have not obtained a public implementation or model weights.

PROBE uses reconstruction-based training data similar to ours. It then adds detector-guided hard samples to broaden the training distribution and to improve generalization. Despite PROBE's use of additional training samples, PRL achieves an average BAcc of 95.67% across the 11 benchmarks, exceeding PROBE's 93.34% by 2.33%. On the most challenging benchmark, Chameleon, PRL achieves 90.01%, outperforming PROBE's 87.16% by 2.85%. These results support PRL's design: exploiting fine-grained correspondence within existing aligned pairs and accounting for contextual shifts yields stronger overall detection performance without additional training images or a detector-guided generation stage. We will explore combining PRL with PROBE-style hard-sample generation to further improve coverage of difficult cases. PGC shows sensitivity to input image format and achieves an average BAcc of 42.23%, with PRL outperforming it on all 11 benchmarks.

The table below reports BAcc (%).

MethodGenImageAIGCDDRCT-2MDDA-COCOEvalGENSynthbusterForenSynthsUnivFDChameleonSynthWildXWildRFAvg.
PROBE97.4791.9999.6891.1589.5698.2386.6690.6987.1697.2296.9293.34
PGC51.9336.0661.1650.014.9845.9816.6429.7853.4452.0862.5142.23
PRL (Ours)96.8397.7599.2794.3999.1497.0092.5794.2290.0193.6097.5995.67

We add discussions of these works to the Related Work section and the quantitative comparison table above to Appendix F of the revised paper.

AuthorsOpenReview ↗

Follow-up on Author Response

Dear Reviewer wHne, thank you again for your thoughtful review. In our rebuttal, we clarified the scope of reference coherence and its role in regularizing detection-related score changes (W1). Following your suggestion, we evaluated direct response supervision without reference subtraction: average BAcc decreases from 95.55% to 92.65%, supporting the reference-based design (W2/Q1). We also discussed the DeepFake limitation with additional AUROC/AP results and threshold analysis (W3), and expanded the discussion of PROBE, PGC and SACG, including quantitative comparisons with PROBE and PGC (Q2). We would appreciate it if you could let us know whether our responses have addressed your concerns.

Reviewer_hpfqOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Prior AI-generated image detection methods use mixed views from aligned real-generated pairs and supervise based on labels defining the source of patch. In vision transformers, self-attention allows patches to interact, which leads to contextual shift: features of a patch moving with its context (i.e. comprising shifts caused by other patches as well); the patch label, therefore, no longer correctly defines the patch features. The proposed method, PRL, improves supervision under such contextual shift by considering two parts: 1) aligned region mixing that creates two mixed views by uniquely mixing patches of real and generated images, and 2) using those two mixed views, perform response learning. In (2), the relative response of source-changed patches is compared against same-source patches (as reference); this removes contextual shift and keeps accurate supervision, unlike labels. Along with this, they also include reference coherence, to maintain stable reference, and area ranking, to associate larger generated area with higher score. With these three aspects, the overall objective leads to an improved, generalizable AI-image detection. The method’s superiority (higher avg. balanced accuracy) over baselines is proven with comparison to 11 baselines and 11 benchmarks.

Critical strengths
  1. The method is carefully constructed and solves a valid issue of prior methods. The authors give proper justification and explanation for each loss term involved in the objective.
  2. Detailed in terms of evaluation: They show the better performance of PRL by comparing against 11 baselines and 11 benchmarks (and also shows results per subset for the benchmarks). Also, they are careful enough to ensure that detectors don’t rely on format bias by bringing synthetic and real images to the same format in case of benchmarks that differed in their formats.
  3. Additionally, the authors also evaluate robustness of PRL to post-processing on a dataset, in addition to the generalizability checks reflected by point 2.
  4. The paper is easy to follow.
  5. The ablations I had in mind are already covered in the main paper and Appendix.
Critical weaknesses

I did not identify weakness as such. However, I have a few questions.

Questions for authors
  1. Since we already get a scalar patch score from the model as shown in Section 3.2 (Patch response), can we not directly use it with sigmoid to get y^\hat{y}, instead of introducing gg?
  2. "If a set in τr\tau^r or τs\tau^s is empty, that term is dropped and the other is used alone." in Section 3.2 (Relative response): how often does this happen, because it directly connects to how much of the contextual shift is addressed?
3 comments
AuthorsOpenReview ↗

Author Response

We thank the reviewer for the careful reading and the positive evaluation. We also thank the reviewer for recognizing our justification of each loss term, the comparison with 11 methods on 11 benchmarks (including the format alignment that removes format bias), and the robustness experiments.

Q1: Since we already get a scalar patch score from the model as shown in Section 3.2 (Patch response), can we not directly use it with sigmoid to get y^\hat{y}, instead of introducing gg?

Indeed, the patch scores can be aggregated and passed through a sigmoid function to obtain an image-level prediction. However, we adopt a separate detection head gg for two reasons: (1) Objective decoupling: The patch scoring head 𝐰 focuses on learning relative responses between aligned views, while gg independently learns image-level discrimination. (2) Adaptive aggregation: Unlike uniform averaging of patch scores, gg learns input-dependent weights to emphasize more informative patches.

Experimental validation. We replace gg with y^=σ(meanpzp)\hat{y} = \sigma(\mathrm{mean}_{p} z_{p}) and keep all other settings unchanged. Using gg achieves higher average BAcc across the 11 benchmarks (+0.93%), with an improvement on the in-the-wild benchmarks (+2.63%). (Diff. = σ(meanp zp)\sigma(\mathrm{mean}_p\, z_p) − gg) :

Detection headGenImageAIGCDetectDRCT-2MDDA-COCOEvalGENSynthbusterForenSynthsUnivFDChameleonSynthWildXWildRFStandardWildAvg.
gg (PRL)96.8397.7599.2794.3999.1497.0092.5794.2290.0193.6097.5996.4093.7495.67
σ(meanp zp)\sigma(\mathrm{mean}_p\, z_p)97.7497.1499.2696.6699.5596.8790.2291.4286.3092.2494.7996.1191.1194.74
Diff.+0.91−0.61−0.01+2.27+0.41−0.13−2.35−2.80−3.71−1.36−2.80−0.29−2.63−0.93

 

Q2: “If a set in τr\tau^{\mathrm{r}} or τs\tau^{\mathrm{s}} is empty, that term is dropped and the other is used alone.” in Section 3.2 (Relative response): how often does this happen, because it directly connects to how much of the contextual shift is addressed?

We thank the reviewer for this question. We use the mask sampler of Appendix A and sample 200k training pairs with two mixed views. τr\tau^{\mathrm{r}} is used only when both Sr,s\mathcal{S}^{\mathrm{r,s}} and its reference Sr,r\mathcal{S}^{\mathrm{r,r}} are non-empty. τs\tau^{\mathrm{s}} is used only when both Ss,r\mathcal{S}^{\mathrm{s,r}} and its reference Ss,s\mathcal{S}^{\mathrm{s,s}} are non-empty. We repeat the sampling with two random seeds, and the results differ by less than 0.05%. As shown in the table below, at least one term is always used during training, and both terms are used in nearly three quarters of the pairs. Thus, reference-based supervision for addressing contextual shifts remains active in every sampled pair, covering both source-change directions in most pairs.

Training pairs with two mixed viewsFrequency
Both τr\tau^{\mathrm{r}} and τs\tau^{\mathrm{s}} used74.4%
At least one term dropped25.6%
At least one term used100.0%
Neither term used0.0%

Thank you for the clarification. The authors' responses have addressed my questions.

Thank you for your positive feedback. We are glad that our responses have addressed your questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Paper Revisions

Based on the reviewers' valuable suggestions, we update our manuscript (modifications are marked in blue) with the following main additions and clarifications:

  1. Comprehensive Threshold-Free Evaluations (Sec. 4.1, Table 2, and Appx. D, Tables 7, 9, 11, 13, 15, 17-19; to GNac): We explicitly clarify the uniform 0.50.5 decision threshold in Sec. 4.1 and add overall and per-subset AUROC and AP comparisons across all benchmarks.
  2. Generalization to Alternative Vision Backbones (Sec. 4.5 and Appx. E, Table 23; to GNac): Alongside our same-backbone comparison with DDA on DINOv3 (Table 22), we add controlled experiments on CLIP ViT-L/14, showing that PRL consistently improves over image-level BCE on a backbone with a different pretraining objective.
  3. Ablation on Direct Response Supervision without Reference Subtraction (Appx. E, Table 21; to wHne): We add an ablation variant that supervises source-changed patch responses without subtracting the reference group responses, empirically verifying that accounting for shared contextual shifts via reference subtraction is essential.
  4. Discussions and Quantitative Comparisons with Concurrent Works (Sec. 2 and Appx. F, Table 24; to wHne): We expand the Related Work section to discuss PROBE (ICML 2026), PGC (ICML 2026), and SACG (CVPR 2026), and add quantitative comparisons in Appx. F, showing PRL's superior performance.
AuthorsOpenReview ↗

General Response and Summary of Discussion

We sincerely thank the Area Chair and all reviewers for their time, careful reading, and constructive feedback. We are encouraged that all reviewers (hpfq, wHne, GNac) gave positive ratings (4: Clear acceptance, 3: Weak acceptance, 3: Weak acceptance), recognizing that our paper not only addresses the underuse of correspondence in aligned pairs, but more critically identifies and resolves contextual shift: when real and synthetic patches are mixed, self-attention couples their features, spreading generation traces into real patches while diluting synthetic ones, which renders per-patch source labels imprecise. To overcome this challenge, we propose PRL, a well-motivated framework that supervises relative patch responses across mixed views, validated through broad evaluations across 11 baselines and 11 standard/in-the-wild benchmarks, local manipulation tasks, and post-processing robustness checks.

During the discussion period, we actively engaged with all reviewers to address their questions and suggestions regarding objective design, simpler correspondence-based baselines, reference subtraction, backbone generalization, evaluation metrics/thresholds, and comparisons with concurrent works. Below is a summary of the discussion.

ReviewerRatingStrengthsQuestions & ConcernsOur Reply
hpfq4: Clear acceptance(1) Carefully constructed method solving a valid issue (contextual shift) with proper justification for each loss term. (2) Detailed evaluation across 11 baselines and 11 benchmarks. (3) Thorough robustness evaluation against post-processing. (4) Clear presentation and comprehensive ablations.(Q1) Why use a separate image-level head? (Q2) How often is a relative-response term dropped?(Q1) The separate head decouples image-level discrimination from patch scoring. Compared with applying a sigmoid to the mean patch score, it raises mean BAcc from 94.74% to 95.67%. (Q2) In 200,000 sampled pairs, at least one term is active in every pair and both are active in 74.4%.
wHne3: Weak acceptance(1) Well-motivated observation on contextual interactions in ViTs. (2) Intuitive and effective relative response formulation using unchanged patches as references. (3) Strong experimental performance across 11 benchmarks, especially on challenging in-the-wild datasets.(W1) Coherence may overconstrain semantically different patches. (W2, Q1) Compare with direct response supervision and explain how subtraction handles varying contextual effects. (W3) Discuss limited generalization on DeepFake face swaps. (Q2) Discuss PROBE, PGC, and SACG.(W1) Coherence regularizes detection-score changes of unchanged patches by source. Mean BAcc is 94.47-95.55% with coherence and 94.05% without it (Fig. 6). (W2, Q1) The reference mean estimates shared effects. Without subtraction, four-benchmark mean BAcc falls from 95.55% to 92.65% (Table 21). (W3) DeepFake BAcc is 55.26% at a threshold of 0.5, with 81.11% AUROC and 80.53% AP. In an exploratory analysis on this subset, BAcc rises to 73.05% at a threshold of 0.32. (Q2) We discuss PROBE, PGC, and SACG (Sec. 2) and compare with PROBE and PGC (Appx. F, Table 24). PRL leads PROBE overall (95.67% vs. 93.34%) and on Chameleon (90.01% vs. 87.16%).
GNac3: Weak acceptance(1) Supervision design effectively accounts for contextual shift with clear visualizations. (2) Better exploitation of fine-grained spatial correspondence within aligned pairs. (3) Broad evaluation across cross-generator, in-the-wild, and local manipulation benchmarks. (4) Well-controlled ablation studies.(W1, Q1) Isolate PRL's supervision gains under matched settings and test them on CLIP. (Q2) Clarify the BAcc threshold and report AUROC and AP.(W1, Q1) With DINOv3, LoRA, and training pairs fixed, PRL improves four-benchmark mean BAcc over image-level BCE from 92.65% to 95.55% (Table 21). PRL averages 95.67% versus 93.20% for DDA with the same setup across 11 benchmarks (Table 22). On CLIP, PRL improves mean BAcc over BCE from 74.43% to 83.90% on four benchmarks (Table 23), and from 72.97% to 84.39% across 11 benchmarks in our reply. (Q2) We use a fixed threshold of 0.5 (Sec. 4.1). PRL leads on both metrics across ten benchmarks with both classes, averaging 98.83% AUROC and 98.55% AP (Table 2, Appx. D).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.