Organizations: Guangdong University of Technology · The Hong Kong University of Science and Technology (Guangzhou) · Southern University of Science and Technology · State Key Laboratory of Public Big Data, Guizhou University
AI-generated image detectors face significant challenges when deployed in real-world environments, particularly when test samples are produced by unseen generators that deviate from the source training distribution. Existing detectors often rely on a single type of evidence, such as semantic representations or low-level generation traces, making them vulnerable when the corresponding cue becomes unreliable. Moreover, adapting detectors to newly emerging generators usually requires target labels or retraining, which is costly and impractical. To address these challenges, we propose Dual-Evidence Disagreement-Constrained Adaptation (DEDCA), a test-time adaptation framework for generalized AI-generated image detection. DEDCA constructs a dual-evidence detector by combining CLIP-based semantic evidence with trace evidence extracted from shuffled entropy and residual statistics, and introduces a reliability-aware gate to perform sample-specific adaptive fusion. Our key idea is to exploit the disagreement between the two evidence streams as an unlabeled adaptation signal, enabling the detector to adjust to target-domain shifts without relying on target labels or blindly trusting a single prediction stream. Experiments show that DEDCA achieves superior performance on the GenImage benchmark by outperforming state-of-the-art AI-generated image detectors and test-time adaptation methods with 92.0% average accuracy.
Figures & tables
Figure 1: Motivation of dual-evidence disagreement for test-time adaptation. (a) Complementary error patterns between semantic and trace evidence. (b) Cross-evidence disagreement as an indicator of unreliable predictions.
Figure 2: Overall framework of DEDCA. The training stage learns a dual-evidence detector with SERT-based trace modeling and DERG-based reliability fusion, while the test-time stage adapts the detector on unlabeled target batches through disagreement-constrained optimization.
Table 3
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Parameter and gate analysis of DEDCA under the leave-ADM setting. (a) Effect of source-domain loss balance, where λced=λrel=λRMA=λa and λsp=0.1 . (b) Effect of test-time adaptation loss balance, where λent=0.5 and λgate=λdrift=λb . (c) Effect of the number of adaptation steps Kstep . (d) Evolution of the average gate weight mˉ during source-domain training.
Method
Color Saturation
Color Contrast
Gaussian Blur
Resize
JPEG
Average
ACC
AP
Δ
ACC
AP
Δ
ACC
AP
Δ
ACC
AP
Δ
ACC
AP
Δ
ACC
AP
Δ
MoLD
89.0
96.8
2.4
88.1
96.0
3.3
86.7
94.8
4.7
87.2
95.2
4.2
86.9
95.5
4.5
87.6
95.7
3.8
T2A-DEDCA
83.9
94.0
4.0
82.5
92.8
5.4
80.6
91.2
7.3
81.0
91.6
6.9
81.7
92.1
6.2
81.9
92.3
6.0
ReservoirTTA-DEDCA
77.5
90.6
3.8
77.0
89.2
4.3
75.1
87.5
6.2
74.9
88.1
6.4
76.4
88.6
4.9
76.2
88.8
5.1
DEDCA (w/o TTA)
85.2
96.1
4.2
83.4
94.8
6.0
81.0
92.6
8.4
81.8
93.2
7.6
82.5
93.9
6.9
82.8
94.1
6.6
DEDCA
90.4
98.0
1.6
88.6
97.0
3.4
86.8
95.9
5.2
86.7
95.6
5.3
88.0
96.7
4.0
88.1
96.6
3.9
Appendix
Table 5: Robustness evaluation under common post-processing operations. Δ denotes the absolute ACC drop compared with the clean average performance of each method.
Figure 4: t-SNE visualization of real and fake feature distributions produced by DEDCA (w/o TTA). Each subplot corresponds to one held-out generator, where blue and red points denote real and fake samples, respectively.
Figure 5: t-SNE visualization of real and fake feature distributions produced by DEDCA. Each subplot corresponds to one held-out generator, where blue and red points denote real and fake samples, respectively.
Figure 6: Multi-generator t-SNE visualization of the feature space produced by DEDCA without test-time adaptation.
Figure 7: Multi-generator t-SNE visualization of the feature space produced by DEDCA with test-time adaptation.
Method
Total Params
Params@TTA
Infer Time
TTA Time
MLEP
1.40M
–
1.42
–
LARE 2
104.31M
–
0.85
–
MoLD
303.71M
–
10.72
–
Xception
20.81M
–
1.57
–
T2A (Xception)
20.81M
54.5K
1.57
11.70
PeTTA (ResNet-50)
25.56M
53.1K
0.65
3.63
Appendix
Table 6: Computational cost comparison. Infer Time denotes the non-TTA inference time. Params@TTA denotes the number of parameters updated during test-time adaptation.
AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.
Detecting AI-generated images (AIGI) remains challenging because detectors often fail to generalize to unseen generators. Although existing methods are trained on large datasets, their performance still degrades when generation settings change, indicating that data scale alone is insufficient and that limited coverage of generative variations during training is a key factor. Studies on generative model editing show that small changes in internal representations can produce diverse and meaningful image variations, many of which are not explored under standard sampling. Leveraging this insight, we propose PROBE (Probing Robustness via Boundary Exploration), a framework that improves detector generalization by actively exploring challenging regions of the generative process. Instead of treating the generator as a fixed data source, PROBE uses the detector as a critic to steer the generator through manifold-level modifications, producing realistic samples that are difficult to classify. These samples expose failure cases that are uncommon under standard data sampling strategies and are used to refine the detector. Experimental results across multiple benchmarks indicate that PROBE enhances generalization to unseen generators, resulting in more generalizable AIGI detection performance. Code and models are available at https://github.com/Amamiya-C/PROBE-AIGI-Detection
Zijie Cao, Weijie Tu, Yao Xiao +3
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Australian National University, Canberra, Australia · Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China +1
AI-generated image detectors deployed in open-world environments inevitably face distribution shifts as new and stronger generative models continue to emerge. Although existing methods improve cross-generator generalization through better representations or training data construction, they typically follow a static train-once-and-deploy paradigm and cannot adapt after deployment. In this work, we study open-set AIGC image detection from a test-time adaptation perspective. We propose Test-Time Curriculum (TTC), a simple and model-agnostic framework that adapts a detector on unlabeled test data through curriculum-based self-training. TTC starts from highly reliable pseudo-labeled samples and progressively incorporates harder yet informative cases, while enforcing class-balanced selection to reduce biased updates under generator shift. To further improve pseudo-label quality, we introduce Cross-Scale Pseudo-Label Refinement, which aggregates complementary evidence across multiple resolutions for more reliable adaptation, and applies noisy-or fusion at inference to strengthen final predictions. In addition, we construct AIGCGuard, a new benchmark containing 3,100 representative real images and 124,000 generated images from 40 of the most advanced open-source and proprietary text-to-image models. Extensive experiments on five benchmarks show that TTC substantially improves overall detection performance under diverse unseen-generator shifts, establishing a practical and effective test-time adaptation framework for open-set generated image detection.
Yiqian Zhang, Zheyuan Gu, Xiangzhao Hao +8
Hangzhou Dianzi University · ERNIE Team, Baidu Inc. · Harbin Institute of Technology (Shenzhen)
The rapid advancement of generative Artificial Intelligence (AI) has introduced significant challenges for reliable AI-generated image detection. Existing detectors often suffer from performance degradation under distribution shifts and when encountering newly emerging generative models. In this work, we propose a data-centric continual adaptation framework for updating detectors in evolving environments. We show that both in-the-wild data and generator-driven data are essential for adapting detectors. We introduce an automated, weakly supervised pipeline for constructing in-the-wild datasets through fact-check article retrieval. Additionally, we demonstrate that incorporating even a small amount of generator-driven data during training enables effective adaptation to newly emerging models, while combining it with in-the-wild data within a continual learning framework enables robust adaptation and mitigates catastrophic forgetting. Extensive experiments on two state-of-the-art detectors show significant improvements of +9.14% and +8% in average accuracy, respectively.