Recent works in AI-generated image detection have shown that careful training data alignment can improve generalization by removing spurious correlations. However, linear probes on frozen DINOv3 representations achieve remarkably strong performance even when trained on misaligned datasets. Motivated by this result, we analyze the underlying rationale and the limits of this generalization. We find that frozen DINOv3 performs well because its decisions rely on features that faithfully represent the space of authentic images. At the same time, its final layer is less effective at capturing the subtle pixel-artifact cues that can be emphasized by aligned training data. We further observe that naively mixing aligned and misaligned data during adaptation improves sensitivity to such cues but at the cost of distorting the pre-trained representation, limiting generalization. To address this issue, we propose Anchor-Regularized Adaptation (ARA). We apply Low-Rank Adaptation to capture pixel-level artifacts while leveraging a frozen anchor classifier to avoid deviations from the original representation structure. This allows the model to exploit pixel-artifact cues without sacrificing generalization. Our method achieves state-of-the-art performance on nine diverse and challenging benchmarks, indicating that ARA enables complementary supervision from misaligned and aligned data for more effective detection.
Figures & tables
Figure 1. Conceptual illustration of the structural trade-off in adapting DINOv3 for AIGI detection. (a) Frozen DINOv3 provides a strong authenticity prior but struggles with fakes containing only subtle pixel-artifacts. (b) Naive fine-tuning can absorb such cues while distorting the pre-trained feature structure, degrading generalization. (c) ARA incorporates new artifact cues while better preserving the pre-trained structure. A three-panel overview contrasting frozen DINOv3, naive fine-tuning, and the proposed ARA method in terms of feature-space structure and forgery detection behavior.
Figure 2. Empirical analysis of frozen DINOv3. (a) Reducing the amount of real training data substantially degrades performance, confirming reliance on a real-centric manifold . (b) Layer-wise probing on aligned data reveals a semantic bottleneck at the final layer. (c) Fusing predictions from intermediate experts compensates for this bottleneck, improving in-the-wild generalization. Three plots showing data-ratio sensitivity, layer-wise probing accuracy, and decision-fusion gains for frozen DINOv3 under aligned and misaligned training regimes.
Figure 3. Explained variance ratios of the top 10 principal components computed from real features. The increasing PC1/PC2 ratio in Cases 2 and 3 indicates stronger anisotropic distortion under backbone adaptation. Explained variance ratio of the top principal components across the compared adaptation regimes.
Figure 4. Overview of the Anchor-Regularized Adaptation (ARA) framework. In Stage 1, a linear anchor head hanchor is trained on frozen DINOv3 final layer features using GenImage and then fixed as a geometric reference. In Stage 2, the DINOv3 backbone is adapted with LoRA on the hybrid dataset, where standard GenImage fakes and AlignedGenImage fakes are both treated as the binary fake class. The resulting features are supervised by both a main classification loss Lmain through hmain and an anchor regularization loss Lanchor through the frozen head hanchor . A pipeline diagram showing Stage 1 anchor construction on frozen DINOv3 features and Stage 2 LoRA adaptation regularized by the frozen anchor during hybrid-data training.
(a) Aligned Fake Ratio
r=0.00
91.5
r=0.25
94.4
r=0.50
96.9
r=0.75
96.5
Table 8. Comparison of balanced accuracy (%) under different ablation settings on the in-the-wild benchmark average.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5. Layer-wise probing on B-Free and DDA also reveals a semantic bottleneck . Layer-wise probing on B-Free and DDA also reveals a semantic bottleneck.
Figure 6. Fusing predictions from intermediate experts compensates for this bottleneck, improving generalization. Fusing predictions from intermediate experts compensates for this bottleneck, improving generalization.
Method
Avg.
Naive LoRA ( λ=0 )
92.7
Real Feature Matching
93.5
Anchor Loss on Real Only
94.3
Anchor Loss on Real and Original Fake
94.8
ARA (Ours)
96.9
Appendix
Table 9. Comparison of internal regularization variants in balanced accuracy (%), averaged over the five in-the-wild benchmarks.
Method
BACC ↑
AUROC ↑
AUPRC ↑
EER ↓
FPR95 ↓
Feature Dist.
89.8
98.5
98.6
5.1
8.7
Output Dist.
92.9
98.3
98.8
4.8
12.6
L2-SP
95.1
98.8
98.7
4.0
3.6
LP-FT
79.2
96.0
97.1
9.3
22.0
WiSE-FT
90.2
98.5
98.7
4.6
7.2
ARA (Ours)
96.9
99.3
99.0
2.7
1.9
Appendix
Table 10. Comparison of regularized adaptation baselines, averaged over the five in-the-wild benchmarks.
Backbone
Linear Probe
Naive LoRA
ARA (Ours)
DINOv3-7B
93.4
92.7
96.9
DINOv3-H+
87.5
84.4
93.2
DINOv3-L
78.6
88.3
91.4
DINOv3-B
73.3
62.8
63.5
Appendix
Table 11. Comparison of balanced accuracy (%) across DINOv3 backbone scales on the in-the-wild benchmark average.
Figure 7. Balanced accuracy (%) under JPEG compression, resizing, and Gaussian blur on Synthbuster. We compare the frozen DINOv3 linear probing baseline with our ARA. Three line plots comparing the frozen DINOv3 linear probing baseline and ARA on Synthbuster under JPEG compression, resizing, and Gaussian blur. The y-axis shows balanced accuracy, and ARA remains consistently above the baseline across perturbation levels.
AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.
Recent advances in AI-Generated Image (AIGI) detection have largely focused on model architectures, loss design, and training strategies, while overlooking a more fundamental factor: how supervision is structured. Existing detectors typically train on benchmark-defined data that implicitly blends semantic and artifact signals, leaving supervision design unexplored. In this paper, we examine supervision structure through controlled experiments and show that such blended supervision is not merely suboptimal but fundamentally counterproductive. Replacing it with disentangled, orthogonal supervision - paired with inductively matched models - suffices to achieve strong cross-generator generalization. We formalize this as Orthogonal Supervision Disentanglement and instantiate it in AlignGemini, a two-branch detector in which a VLM receives purely semantic supervision and a vision expert receives purely artifact supervision. With disentangled supervision coupled with inductively matched models, even the simplest training recipe achieves a significant improvement of 9.5% across five in-the-wild benchmarks - trained on images from only two generators, without complex training strategies or diverse generator coverage. We further introduce AIGI-Now, a diagnostic benchmark covering eight contemporary generators, including five closed-source commercial models, with disentangled semantic and pixel evaluation subsets for fine-grained analysis of detector capabilities. Our code is publicly available at https://github.com/GzZning/Align-Don-t-Blend.
Ruoxin Chen, Jiahui Gao, Kaiqing Lin +5
Tencent Youtu Lab · East China University of Science and Technology · Shenzhen University +1
AI-generated image detectors generalize poorly when their training and test images originate from different generators or datasets. Despite the rich spatial representations produced by vision foundation models like DINO, existing detectors typically classify images using only the globally aggregated CLS token. We hypothesize that globally aggregating DINO features into a single CLS token obscures spatially distributed generation traces. To test this hypothesis, we introduce PatchHead, a lightweight spatial aggregation head that preserves the two-dimensional organization of DINO patch tokens and integrates evidence across neighboring regions. During training, we freeze the pretrained DINO backbone and optimize only the inserted LoRA adapters, PatchHead, and auxiliary projection head. Across nine cross-dataset benchmarks spanning manually curated and in-the-wild settings, PatchHead ranks first on seven datasets and second on the remaining two. It improves the strongest prior method from 91.6% to 94.6% in average balanced accuracy (+3.0 points) and raises the worst-case accuracy from 82.4% to 89.4% (+6.9 points), while introducing only 8.6% more trainable parameters and 0.08% additional FLOPs. Further qualitative analysis suggests that PatchHead (i) reduces class-conditional domain discrepancy, and (ii) redirects the representation from content-dominated saliency toward spatially distributed authenticity evidence. Together, these observations provide a representation-level account of why spatial patch aggregation transfers more reliably across generators and datasets than a single CLS-based global representation. Our code and models will be made available upon acceptance.
Shengbo Qi, Hongyi Fang, Benjia Zhou +1
Beijing Institute of Technology, Zhuhai · Shenzhen University
Generalization remains a critical bottleneck in AI-generated image detection. Because many modern generators are proprietary or adversarially modified, existing detectors overfit to the low-level textural patterns of accessible training data, resulting in severe failures on unseen domains. Conventional regularization techniques (e.g., L1/L2 norms, Dropout) apply indiscriminate parametric constraints and fail to provide the domain-invariant structure necessary for cross-generator robustness. To address this, we propose Feature-Augmented Implicit Regularization (FAIR). FAIR introduces an orthogonal, macro-structural prior, specifically, Scene Composition Structure (SCS), during training to geometrically constrain the model's optimization trajectory. By augmenting the primary feature space with domain-invariant SCS features, FAIR explicitly penalizes texture-biased shortcut learning. Crucially, this structural prior is entirely discarded at inference, yielding a smoothed, generalized decision boundary with zero architectural or computational overhead. Extensive evaluations across five massive benchmarks demonstrate that integrating FAIR into state-of-the-art detectors significantly improves cross-generator generalization, boosting accuracy by up to 8.04% and establishing new state-of-the-art robustness in zero-shot transfer scenarios.
Md Redwanul Haque, Manzur Murshed, Manoranjan Paul +1
Deakin University, Burwood, VIC, Australia · Charles Sturt University, Bathurst, NSW, Australia