Leveraging Arbitrary Data Sources for AI-Generated Image Detection Without Sacrificing Generalization

Qinghui He; Haifeng Zhang; Xiuli Bi; Bo Liu; Chi-Man Pun; Bin Xiao

cs.CVFeb 28, 2026

Leveraging Arbitrary Data Sources for AI-Generated Image Detection Without Sacrificing Generalization

Authors: Qinghui He, Haifeng Zhang, Xiuli Bi, Bo Liu, Chi-Man Pun, Bin Xiao

Organizations: Chongqing Key Laboratory of Image Cognition, School of Computer Science and Technology, Chongqing University of Posts and Telecommunications, Chongqing, China · Chongqing Key Laboratory of Image Cognition, School of Artificial Intelligence, Chongqing University of Posts and Telecommunications, Chongqing, China · Chongqing Key Laboratory of Image Cognition, Chongqing University of Posts and Telecommunications, Chongqing, China

Abstract

The accelerating advancement of generative models has introduced new challenges for detecting AI-generated images, especially in real-world scenarios where novel generation techniques emerge rapidly. Existing learning paradigms are likely to make classifiers data-dependent, resulting in narrow decision margins and, consequently, limited generalization ability to unseen generative models. We observe that both real and generated images intend to form clustered low-dimensional manifolds within high-level feature spaces extracted by pre-trained visual encoders. Building on this observation, we propose a single-class attribution modeling framework that first amplifies the intrinsic differences between real and generated images by constructing a compact attribution space from any single-class training set, either composed of real images or generated ones, and then establishes a more stable decision boundary upon the enlarged separation. This process enhances class distinction and mitigates the reliance on generator-specific artifacts, thereby improving cross-model generalization. Extensive experiments show that our method generalizes well across various unseen generative models, outperforming existing detectors by as much as 7.21% in accuracy and 7.20% in cross-model generalization.

Figures & tables

Figure 1: (a) Previous methods overfit to training distributions with small margins, while our attribution space yields larger margins and stronger cross-model generalization. (b) The overall performance of our method across the entire dataset compared to the SOTA methods. (c) The cross-model generalization performance of our method compared to the SOTA methods.

Figure 2: Visualization of CLIP features and attribution deviations. (a) UMAP visualization of CLIP features for real and generated images. (b) UMAP visualization of attribution deviations under $\mathcal{M}_{r}$ . (c) UMAP visualization of attribution deviations under $\mathcal{M}_{g}$ . (d) Violin plots of attribution deviations using $\mathcal{M}_{r}$ (left) and $\mathcal{M}_{g}$ (right). (e) Quantitative comparison of class separability using inter-class distance, intra-class variance, and Fisher’s Ratio.

Figure 3: Overview of the proposed framework. Given real images $x_{r}$ and generated images $x_{g}$ , a frozen pre-trained model $f$ first extracts high-level semantic features. The proposed Attribution Consistency Module then constructs an attribution space $\mathcal{A}$ using samples from a single class (either real or generated) to learn class-consistent representations. Features from the opposite class that deviate from this learned manifold are treated as attribution shifts. A lightweight classifier $\mathcal{C}$ is finally applied to distinguish real and generated images based on their attribution-space representations.

Figure 4: Performance comparison of SOTA detectors (a-g) and our method (h) trained on SD1.4, evaluated on 15 generative models using ACC , AP , F1 , Real ACC , and Fake ACC . Our approach achieves balanced real-fake accuracy and superior cross-model generalization.

Methods	Generative Adversarial Networks						Others		Diffusion Models							AP
Methods	Pro- GAN	Big- GAN	Cycle- GAN	Star- GAN	Style- GAN	Gau- GAN	SAN	Deep- Fake	ADM	GLIDE	SD1.4	SD1.5	VQDM	Midj	Wukong	avg.
CNNDet [ 36 ]	100.00	84.52	93.46	98.15	99.54	89.49	57.67	89.03	71.05	66.09	56.88	57.25	61.90	55.89	52.85	75.58
LNP [ 19 ]	99.98	95.20	98.14	100.00	99.56	83.25	45.29	57.72	99.25	96.45	89.77	89.30	95.54	94.39	91.16	89.00
LGrad [ 33 ]	100.00	90.76	94.02	99.98	99.81	79.29	45.10	71.71	71.84	75.96	70.91	71.73	70.23	71.43	66.51	78.62
DIRE [ 37 ]	69.76	51.92	64.76	67.98	71.40	88.18	54.89	67.90	44.99	58.11	47.46	47.06	85.18	55.55	51.25	61.76
UnivFD [ 25 ]	100.00	99.23	99.76	99.01	97.88	99.98	77.53	81.22	86.48	83.24	86.00	85.17	96.50	73.35	90.89	90.45

Table 1: The AP values of various methods trained on GAN source images. Bold values indicate the best performance, and underlined values indicate the second best.

Methods	Generative Adversarial Networks						Others		Diffusion Models							ACC
Methods	Pro- GAN	Big- GAN	Cycle- GAN	Star- GAN	Style- GAN	Gau- GAN	SAN	Deep- Fake	ADM	GLIDE	SD1.4	SD1.5	VQDM	Midj	Wukong	avg.
CNNDet [ 36 ]	99.99	70.08	85.16	91.65	85.73	78.94	50.91	53.49	58.73	54.96	51.57	51.78	53.64	52.61	50.23	65.96
LNP [ 19 ]	99.09	88.05	91.63	100.00	95.89	79.70	43.61	53.99	94.53	87.84	68.58	68.06	77.89	79.92	73.89	80.18
LGrad [ 33 ]	99.83	82.90	85.28	99.60	94.66	72.49	44.47	56.42	66.96	62.75	67.20	67.91	67.03	65.69	63.67	73.12
DIRE [ 37 ]	59.18	51.68	60.75	62.11	57.85	82.69	48.32	60.15	44.69	59.14	47.53	47.02	73.83	53.13	50.09	57.21
UnivFD [ 25 ]	99.85	95.08	98.18	93.75	85.37	99.42	57.08	67.38	67.01	62.43	64.20	64.01	85.68	56.18	71.39	77.80

Table 2: The ACC values of various methods trained on GAN source images. Symbols follow the same notation as Table 1 .

Figure 5: Scatter plots of per-sample logits. The X-axis and Y-axis represent the outputs of UnivFD and our method, respectively. Decision boundaries of UnivFD (purple) and ours (red) are shown. More results are in Appendix.

Sources	GANs		DMs		Others		avg.
Sources	AP	ACC	AP	ACC	AP	ACC	AP	ACC
Real (GAN)	99.29	98.04	91.67	88.32	87.44	84.45	94.15	91.69
Real (DM)	97.70	94.01	99.20	96.20	86.27	78.60	96.87	92.98
ProGAN	99.18	98.09	90.96	87.78	89.44	84.06	94.05	91.41
ADM	90.89	88.28	96.64	94.52	87.39	85.64	93.11	90.84
SD1.4	98.19	93.08	99.06	96.46	86.21	78.14	97.00	92.66

Table 3: Performance comparison of models constructed with different attribution sources. Real (GAN) and Real (DM) indicate models built using real images from the ProGAN and SD1.4 datasets, respectively. Others use generated images only. More results are in Appendix.

Figure 6: Performance under limited training data. We train the full pipeline using 0.05%, 0.1%, 1%, 10%, and 100% of the ProGAN dataset (720K images in total). (a,b) Results on $\mathcal{M}_{g}$ , and (c,d) results on $\mathcal{M}_{r}$ .

Figure 7: Robustness evaluation under perturbations with the top row showing JPEG compression and the bottom row showing Gaussian blur.

AI-Generated Image Detection

AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.

See all 194 papers

FAIR: Feature-Augmented Implicit Regularization for AI-generated Fake Image Detection
Jul 24, 2026Md Redwanul Haque, Manzur Murshed, Manoranjan Paul +1Domain Generalization AI-Generated Image Detection
Deakin University, Burwood, VIC, Australia · Charles Sturt University, Bathurst, NSW, Australia
Efficient Zero-Shot AI-Generated Image Detection
Mar 23, 2026Ryosuke Sonoda, Ramya SrinivasanAI-Generated Image Detection
Fujitsu Ltd., Japan · Fujitsu Research of America, Inc., USA
When Detectors Forget Forensics: Blocking Semantic Shortcuts for Generalizable AI-Generated Image Detection
Mar 10, 2026Chao Shuai, Shaojing Fan, Chenlin Zou +6AI-Generated Image Detection Shortcut Learning
State Key Laboratory of Blockchain and Data Security, Zhejiang University, Hangzhou, China · National University of Singapore, Singapore · Chongqing University of Posts and Telecommunications, Chongqing, China

Jul 24, 2026J/K move · Enter open · S save

FAIR: Feature-Augmented Implicit Regularization for AI-generated Fake Image Detection

Md Redwanul Haque, Manzur Murshed, Manoranjan Paul, Tsz-Kwan Lee

Deakin University, Burwood, VIC, Australia · Charles Sturt University, Bathurst, NSW, Australia

Domain Generalization AI-Generated Image Detection Implicit Regularization

Open paper AskPDF arXiv

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.,

L_1

L_2

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.

Abstract

AI-Generated Image Detection

Explore similar work

FAIR: Feature-Augmented Implicit Regularization for AI-generated Fake Image Detection

Efficient Zero-Shot AI-Generated Image Detection

When Detectors Forget Forensics: Blocking Semantic Shortcuts for Generalizable AI-Generated Image Detection