- 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
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 indicating the likelihood that an image is synthetic. We use a fixed threshold of 0.5, classifying an image as synthetic if . 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 (%):
| Method | GenImage | AIGCDetect | DRCT-2M | DDA-COCO | Synthbuster | ForenSynths | UnivFD | Chameleon | SynthWildX | WildRF | Avg. |
|---|---|---|---|---|---|---|---|---|---|---|---|
| UnivFD | 83.81 | 87.95 | 83.48 | 68.41 | 53.86 | 92.05 | 92.68 | 55.37 | 46.13 | 53.08 | 71.68 |
| FatFormer | 63.94 | 76.60 | 31.23 | 48.66 | 51.58 | 95.82 | 92.33 | 58.60 | 58.15 | 72.75 | 64.97 |
| NPR | 38.72 | 34.38 | 45.22 | 48.90 | 38.03 | 23.77 | 26.67 | 32.49 | 37.84 | 60.78 | 38.68 |
| DRCT | 97.54 | 76.83 | 99.29 | 80.22 | 74.69 | 55.26 | 69.88 | 74.64 | 82.46 | 80.81 | 79.16 |
| C2P-CLIP | 81.16 | 84.83 | 82.32 | 56.60 | 62.44 | 97.03 | 95.97 | 39.97 | 54.90 | 60.21 | 71.55 |
| AIDE | 84.75 | 69.70 | 89.04 | 52.98 | 64.78 | 45.05 | 57.93 | 72.51 | 65.71 | 68.32 | 67.08 |
| AlignedForensics | 97.51 | 80.58 | 99.87 | 97.51 | 83.06 | 64.98 | 74.78 | 84.74 | 85.10 | 86.22 | 85.43 |
| Effort | 95.36 | 86.53 | 85.72 | 57.06 | 87.43 | 67.57 | 77.53 | 68.77 | 63.70 | 79.13 | 76.88 |
| SAFE | 53.69 | 46.94 | 45.40 | 50.00 | 59.59 | 39.61 | 41.45 | 57.13 | 49.13 | 62.63 | 50.56 |
| DDA | 98.70 | 96.77 | 99.89 | 98.75 | 95.67 | 92.27 | 92.70 | 91.94 | 96.32 | 96.14 | 95.92 |
| GAPL | 99.18 | 99.37 | 99.60 | 95.72 | 96.57 | 96.42 | 97.51 | 77.02 | 93.74 | 94.64 | 94.98 |
| PRL (Ours) | 99.71 | 99.80 | 99.97 | 99.19 | 99.68 | 97.86 | 98.47 | 95.51 | 98.48 | 99.63 | 98.83 |
AP (%):
| Method | GenImage | AIGCDetect | DRCT-2M | DDA-COCO | Synthbuster | ForenSynths | UnivFD | Chameleon | SynthWildX | WildRF | Avg. |
|---|---|---|---|---|---|---|---|---|---|---|---|
| UnivFD | 82.54 | 87.15 | 82.65 | 67.50 | 60.29 | 91.78 | 92.68 | 47.68 | 52.11 | 57.39 | 72.18 |
| FatFormer | 68.03 | 79.75 | 43.40 | 52.26 | 55.81 | 96.39 | 93.68 | 57.80 | 58.07 | 73.89 | 67.91 |
| NPR | 42.93 | 43.13 | 47.03 | 47.53 | 44.80 | 40.90 | 41.85 | 39.64 | 43.70 | 65.28 | 45.68 |
| DRCT | 96.65 | 76.77 | 99.44 | 83.75 | 77.19 | 55.50 | 69.44 | 71.00 | 83.78 | 84.66 | 79.82 |
| C2P-CLIP | 82.22 | 85.87 | 82.27 | 58.12 | 71.32 | 97.57 | 96.63 | 40.19 | 56.00 | 64.25 | 73.44 |
| AIDE | 86.27 | 69.66 | 90.38 | 55.58 | 67.11 | 48.27 | 60.36 | 65.36 | 65.02 | 72.04 | 68.00 |
| AlignedForensics | 96.09 | 81.22 | 99.89 | 97.28 | 81.81 | 66.23 | 74.84 | 82.85 | 87.24 | 90.13 | 85.76 |
| Effort | 93.65 | 82.90 | 80.13 | 55.81 | 87.67 | 65.36 | 76.64 | 58.57 | 61.17 | 81.40 | 74.33 |
| SAFE | 56.13 | 50.10 | 48.33 | 52.36 | 57.35 | 45.74 | 46.86 | 50.64 | 49.78 | 66.46 | 52.38 |
| DDA | 98.48 | 96.45 | 99.90 | 98.82 | 95.85 | 91.27 | 92.61 | 91.20 | 96.39 | 96.76 | 95.77 |
| GAPL | 99.15 | 99.34 | 99.61 | 95.73 | 96.78 | 96.21 | 97.36 | 74.83 | 94.34 | 94.96 | 94.83 |
| PRL (Ours) | 99.64 | 99.79 | 99.97 | 99.25 | 99.55 | 97.76 | 98.41 | 93.48 | 98.10 | 99.59 | 98.55 |
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).
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.
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.
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:
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)
We add the results on the CLIP backbone to the appendix of the revised paper (Appx. E; Table 23).