- Rating
- 2
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes Adaptive Evidence Selection (AES) for generalizable AI-generated image detection, aiming to improve generalization to unseen generators. It addresses two limitations of many existing vision foundation model–based detectors that rely on the [CLS] token: semantic shortcuts arising from image content and over-reliance on a few salient patches. AES first extracts patch features and groups them using K-means clustering. Cluster-centered Bilinear Aggregation (CBA) then subtracts each cluster’s mean and models low-rank second-order statistics of the residuals to suppress shared semantic content and capture subtle within-cluster variations. Subsequently, Determinantal Evidence Selection (DES) combines relevance scores derived from learnable class prototypes with a determinant-based measure of feature diversity to select compact, complementary evidence separately for the real and generated classes. Experiments across multiple benchmarks support the effectiveness of the framework and demonstrate improved cross-generator generalization.
Critical strengths
1.The motivation is insightful and clearly articulated. The paper identifies two limitations of global representations used by many existing VFM-based detectors—semantic shortcuts and concentrated reliance on a few salient patches—and introduces CBA and DES to address them, respectively. This provides a coherent connection between the motivation and the proposed framework. 2.The experimental evaluation is broad, and the results are competitive. Experiments cover cross-generator and cross-dataset generalization, including an in-the-wild benchmark. Component ablations show that CBA and DES each improve performance, with further gains when combined. The representation analyses provide additional empirical support, while the appendix derives the gradients of the auxiliary prototype loss to help explain its optimization behavior.
Critical weaknesses
1.The use of fixed cluster and subset sizes may limit adaptability to individual images. Although the paper studies different values of K and S, image content varies in complexity, and the amount of informative forensic evidence may also differ across images. A fixed number of clusters and selected descriptors may therefore be suboptimal for some images. 2.The connection between the proposed operations and their intended forensic effects needs stronger justification. Cluster centering does not necessarily disentangle semantic content from generation-related traces, and diversity in feature directions does not necessarily imply complementary forensic evidence. Further analysis would help establish when these assumptions hold. 3.Robustness to common image post-processing operations remains insufficiently evaluated. Although the paper includes an in-the-wild evaluation on Chameleon, it does not systematically assess robustness to JPEG compression, resizing, or blurring. This leaves uncertainty about AES’s reliability under common image degradations encountered in practical deployment.
Questions for authors
1.Could the authors evaluate AES under different levels of JPEG compression? This would help assess its robustness to compression artifacts commonly encountered in real-world images. 2.Does AES generalize to vision foundation models beyond CLIP? Evaluating the framework with a backbone such as DINOv2 would help clarify whether its benefits extend to other pretrained representations or depend on CLIP-specific feature properties.
2 comments
Response to Reviewer Xp92 — Part 2/2**
W2. DES and complementary evidence.
Response: We agree that feature-direction diversity alone does not guarantee complementary forensic evidence. DES combines class relevance with feature diversity, following the quality–diversity principle of determinantal point processes [1]: relevance favors discriminative cluster features, while diversity discourages the selection of similar features.
To examine the effect of diversity, we compare the cosine similarity distributions of the two selected cluster features under full AES and relevance-only selection. After fitting kernel density estimates, the mean cosine similarity decreases from 0.6351 to 0.4265 with diversity (Appendix Fig. 14), indicating that the selected features are less similar.
We also compare detection performance under matched training on GenImage (§4.4, Table 9):
| Selection | mACC (%) |
|---|---|
| Full AES | 98.03 |
| Relevance only | 97.16 |
| Diversity only | 94.85 |
Full AES improves mACC by 0.87 and 3.18 percentage points over relevance-only and diversity-only selection, respectively. Together, these results support combining class relevance and diversity to select less similar features that improve detection.
[1] Kulesza and Taskar. Determinantal Point Processes for Machine Learning. Foundations and Trends in Machine Learning, 2012.
W3 / Q1. JPEG compression, blur, and resizing.
Response: We evaluated AES under JPEG compression, Gaussian blur, and resizing on GenImage. JPEG results for the CLIP configuration are:
| JPEG condition | Clean | Q100 | Q90 | Q80 | Q70 | Q60 |
|---|---|---|---|---|---|---|
| mACC (%) | 97.55 | 96.19 | 93.56 | 90.34 | 87.43 | 84.60 |
Blur and resize results are:
| Operation | Level | mACC (%) |
|---|---|---|
| Gaussian blur | σ=0.5 / 1 / 2 | 95.20 / 90.35 / 88.70 |
| Resize | ×0.5 / ×0.75 / ×1.5 | 88.20 / 94.10 / 91.80 |
These results are reported in the revised paper and appendix.
Q2. Backbones beyond CLIP.
Response: We also report AES results with DINOv2 alongside CLIP on three benchmarks (mACC, %):
| Benchmark | AES (CLIP) | AES (DINOv2) |
|---|---|---|
| GenImage | 98.03 | 96.97 |
| UFD (19 subsets) | 94.97 | 93.46 |
| DRCT-2M (16 domains) | 95.09 | 91.40 |
These results show that AES can also be applied to a visual backbone beyond CLIP.
We hope these results address your concerns.
Response to Reviewer Xp92 — Part 1/2
Thanks for your constructive feedback. Changes are marked in red.
W1: Fixed K and s.
R1: We tested image-dependent cluster and subset sizes on GenImage (§4.4, Table 12; Appendix A.3.1). For adaptive K, we select K ∈ {4, 6, 8, 10} for each training image using the highest mean silhouette score from K-means clustering. For adaptive s, we evaluate s ∈ {1, 2, 3, 4} at inference and select the value whose generated-class probability is farthest from 0.5. Joint adaptation combines these two strategies.
Adaptive K improves mACC by only 0.09 percentage points, while adaptive s and joint adaptation reduce it by 0.18 and 0.49 points. We therefore retain fixed K=8 and s=2 as the default.
Fixed counts do not fix the number of selected patch tokens. As you note, image complexity and the amount of informative forensic evidence may vary across images. Although K and s are fixed, the number of patches in each selected cluster can vary with the image. On GenImage, the reported interquartile range of cluster sizes is 23.71–39.50 patches. With s=2, the real and generated branches select 51.8 and 55.0 patches on average, respectively; their union contains 77.6 distinct patches on average, with an interquartile range of 52–101 patches (Appendix B.1, Fig. 9). Thus, fixed K and s still allow the amount of selected image evidence to vary across images.
W2: Cluster-centered residual features as forensic evidence.
R2: We agree that additional experiments are needed to support this claim. In the revised paper (§4.4 and Appendix B.2), we compare CBA (second-order statistics after cluster centering) with three alternative ways to aggregate features within each cluster: direct cluster means, uncentered second-order statistics, and second-order statistics after image centering.
Semantic probe. We fit a semantic probe to predict the donor image category. CBA reaches chance-level accuracy, substantially lower than the alternatives.
t-SNE. In Appendix B.2, Figure 10 and Table 14, we divide the 20 categories into four groups of five and measure category structure in the original feature space and after t-SNE. Visually, CBA features from the same category are more dispersed, with greater mixing between categories. We use normalized neighborhood entropy to measure category mixing: it ranges from 0 to 1, with higher values indicating greater mixing.
CBA reaches 5.00% accuracy on the 20-class semantic probe, matching the 5% chance level. In the five-class t-SNE analysis, its same-category neighbor fraction is 24.77%, close to the 20% chance level and lower than for the other aggregation methods. Together, these results support the effectiveness of our approach to suppressing semantic information: CBA features contain less object-category information while achieving better detection performance than the compared aggregation methods.