Learning Continuous Source Responses For Generalizable AI-Generated Image Detection

Manni Cui; Ruiqi Liu; Zijian Yu; Hao Tan; Zibo Wei; Zian Wang; Ziheng Qin; Huijia Zhu

cs.CVSep 13, 2026

Learning Continuous Source Responses For Generalizable AI-Generated Image Detection

Authors: Manni Cui, Ruiqi Liu, Zijian Yu, Hao Tan, Zibo Wei, Zian Wang, Ziheng Qin, Huijia Zhu,

Organizations: Ant Group · Huazhong University of Science and Technology · Institute of Automation, Chinese Academy of Sciences · Jilin University

Abstract

Advances in image generation have made synthetic images increasingly difficult to distinguish from real photographs, raising concerns about the trustworthiness of visual media. Existing AI-generated image detectors often perform well on in-domain data, but their robustness and cross-generator generalization remain limited. These limitations are commonly attributed to overfitting to shortcut cues. Although many methods seek to suppress shortcut learning, most retain binary classification as the training task without reconsidering how the task itself shapes the learned representations. We introduce CuRe, a framework for learning Continuous Source Responses that revisits authenticity detection from the perspective of the training task. CuRe reformulates backbone adaptation as regression of real-generated mixing ratios, providing finer supervision that encourages the model to capture authenticity-related variation beyond binary endpoint separation. We further select a compact source-response subspace to suppress nuisance variation and limit the final classifier's access to potential shortcut cues. Across ten public benchmarks, CuRe achieves an average balanced accuracy of 89.7%, exceeding the second-best method by 5.2 percentage points. Further experiments demonstrate consistent generalization gains across visual backbones and strong robustness to common image degradations. Code is available at https://github.com/manic-cui/CuRe

Figures & tables

Figure 1: Comparison of binary and continuous source supervision. Left: binary labels provide only coarse endpoint supervision, allowing models to learn decision boundaries based on different shortcuts (dashed curves), potentially compromising robustness and generalization. Right: CuRe constructs continuous transitions from real ( $\alpha=0$ ) to generated ( $\alpha=1$ ) images and predicts the mixing ratio $\alpha$ , providing finer-grained supervision that encourages the encoder to distinguish intermediate states and become more sensitive to changes in image authenticity.

Figure 2: Representation analysis using PE Core L/14 on GenImage. Frozen and binary-trained features retain pronounced semantic structure, and binary training additionally separates the two JPEG qualities. In contrast, continuous source-change training yields greater mixing across semantic and format groups while producing clearer separation between real and generated samples.

Figure 3: Overview of the CuRe framework. Stage 1 (left) adapts the backbone with LoRA to predict the mixing ratio of real–generated image interpolations, encouraging sensitivity to continuous source changes. Stage 2 (upper right) uses the frozen adapted encoder and real–generated trajectories to extract a compact subspace that preserves source responses while suppressing content and mixing-related variation.

Method	AIGCDetect	AIGIBench	UnivFD	EvalGEN	SynthWildX	WildRF	BFree-On.	RRDataset	RealChain	Chameleon	Overall Avg.
	Standard Benchmarks				In-the-Wild Benchmarks
NPR (CVPR’24) ( Tan et al., 2024 )	63.5	63.9	53.5	86.3	47.4	51.9	46.7	49.3	50.7	54.9	56.8
UnivFD (CVPR’23) ( Ojha et al., 2023 )	67.4	74.7	50.8	94.5	61.6	57.3	64.9	61.2	66.1	62.1	66.1
FatFormer (CVPR’24) ( Liu et al., 2024 )	61.3	63.4	52.6	65.1	52.6	57.7	60.7	50.3	50.0	50.6	56.4
C2P-CLIP (AAAI’25) ( Tan et al., 2025 )	78.4	69.6	91.6	66.4	55.3	59.5	50.0	52.7	50.2	50.7	62.4
DRCT (ICML’24) ( Chen et al., 2024 )	68.3	73.4	56.7	75.4	74.6	73.8	73.8	60.5	64.9	68.5	69.0

Table 1: Overall comparison across AIGI detection benchmarks. All entries are in % and all accuracy values are balanced accuracy. Each benchmark score is the unweighted macro-average over its evaluated subsets when applicable. Best and second-best average results are marked in bold and underlined, respectively.

GAN-based Noise-to-Image Generation
Subset Subset Subset	NPR NPR NPR		UnivFD UnivFD UnivFD		FatFormer FatFormer FatFormer		C2P-CLIP C2P-CLIP C2P-CLIP		DRCT DRCT DRCT		SAFE SAFE SAFE		DDA DDA DDA		GAPL GAPL GAPL		DGS-Net DGS-Net DGS-Net		CuRe CuRe CuRe
Subset Subset Subset	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC
ProGAN	50.4	50.9	52.7	55.1	58.0	84.7	98.5	99.8	54.5	61.5	51.6	53.1	85.3	98.5	99.7	100.0	94.9	99.0	95.2	99.8
R3GAN	62.5	67.5	71.5	80.5	50.3	78.3	83.6	93.9	49.0	40.7	63.0	54.6	95.8	99.1	96.2	99.8	72.4	85.0	96.1	99.0
StyleGAN3	62.6	77.4	65.0	72.5	78.1	92.8	93.1	98.8	52.2	59.5	53.4	62.2	63.3	77.1	97.0	99.7	61.4	79.0	92.6	97.9
StyleGAN-XL	68.4	72.9	75.9	84.1	64.2	86.7	91.3	97.7	49.0	46.1	64.9	79.9	48.8	43.3	95.3	99.0	53.7	72.6	96.9	99.7

Table 2: Subset-level results on AIGIBench. Each method reports B.Acc./AUC (%). Best and second-best results in the Average row are marked in bold and underlined, respectively.

Figure 4: Cross-backbone comparison under different LoRA ranks on AIGIBench. Within each backbone, paired bars compare binary supervision (blue) and continuous-ratio supervision (teal with diagonal hatching) at $r\in\{16,32,64\}$ . Bars report ROC-AUC.

Subset Subset Subset	NPR NPR NPR		UnivFD UnivFD UnivFD		FatFormer FatFormer FatFormer		C2P-CLIP C2P-CLIP C2P-CLIP		DRCT DRCT DRCT		SAFE SAFE SAFE		DDA DDA DDA		GAPL GAPL GAPL		DGS-Net DGS-Net DGS-Net		CuRe CuRe CuRe
Subset Subset Subset	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC
SynthWildX
DALL-E 3	41.1	33.0	63.7	71.8	52.2	65.9	54.9	66.7	85.1	91.8	58.1	57.0	92.4	97.4	87.2	94.3	49.4	51.9	94.6	97.5
Firefly	49.5	46.9	59.0	63.7	52.6	51.8	59.1	78.8	59.3	67.7	53.3	54.8	86.8	93.7	87.4	94.9	51.9	67.2	88.6	94.5
Midjourney v5	51.6	49.0	62.2	70.2	53.2	62.3	51.8	63.4	79.3	88.6	56.5	58.4	92.8	97.9	84.6	92.6	49.5	49.1	94.9	97.7
Average	47.4	43.0	61.6	68.6	52.6	60.0	55.3	69.6	74.6	82.7	56.0	56.7	90.7	96.3	86.4	93.9	50.3	56.1	92.7	96.6

Table 3: Subset-level results on SynthWildX and WildRF. Each method reports B.Acc./AUC (%). SynthWildX pairs each generated subset with the shared real set. Best and second-best results in the Average rows are marked in bold and underlined, respectively.

Table 8

Figure 5: Robustness of CuRe and four strong comparison methods on AIGIBench (BAcc). Each panel varies one post-processing factor while holding the other two at their clean settings.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

GAN-based Generation
Subset Subset Subset	NPR NPR NPR		UnivFD UnivFD UnivFD		FatFormer FatFormer FatFormer		C2P-CLIP C2P-CLIP C2P-CLIP		DRCT DRCT DRCT		SAFE SAFE SAFE		DDA DDA DDA		GAPL GAPL GAPL		DGS-Net DGS-Net DGS-Net		CuRe CuRe CuRe
Subset Subset Subset	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC
BigGAN	45.9	45.3	49.5	55.3	50.2	68.6	98.3	99.9	53.4	61.1	50.0	44.9	87.2	94.8	98.0	100.0	88.3	94.7	93.5	99.7
CycleGAN	70.2	84.7	53.0	60.5	55.5	92.9	98.9	100.0	48.1	51.8	55.7	60.9	71.3	84.9	97.3	99.6	96.7	99.3	70.4	98.6
GauGAN	47.0	46.5	52.1	68.5	50.9	84.3	98.6	99.8	50.1	43.6	49.5	43.8	88.8	95.9	99.4	100.0	88.4	94.4	94.2	99.8
ProGAN	50.4	50.9	52.7	55.1	58.0	84.7	98.5	99.8	54.5	61.5	51.6	53.1	85.3	98.5	99.7	100.0	94.9	99.0	95.2	99.8

Appendix

Table 6: Subset-level results on AIGCDetectBenchmark. Each method reports B.Acc./AUC (%). Best and second-best results in the Average row are marked in bold and underlined, respectively.

Subset Subset Subset	NPR NPR NPR		UnivFD UnivFD UnivFD		FatFormer FatFormer FatFormer		C2P-CLIP C2P-CLIP C2P-CLIP		DRCT DRCT DRCT		SAFE SAFE SAFE		DDA DDA DDA		GAPL GAPL GAPL		DGS-Net DGS-Net DGS-Net		CuRe CuRe CuRe
Subset Subset Subset	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC	B.Acc.	AUC
BigGAN	45.9	45.3	49.5	55.3	50.2	68.6	98.3	99.9	53.4	61.1	50.0	44.9	87.2	94.8	98.0	100.0	88.1	94.5	93.5	99.7
CRN	46.3	23.6	46.7	42.4	49.7	55.9	91.3	98.8	50.3	47.2	51.4	60.5	78.6	89.8	89.8	97.1	61.3	70.1	72.4	84.7
CycleGAN	70.2	84.7	53.0	60.5	55.5	92.9	98.9	100.0	48.1	51.8	55.7	60.9	71.3	84.9	97.3	99.6	96.3	99.1	70.4	98.6
Deepfake	52.6	53.3	54.4	63.8	57.3	75.7	92.6	97.2	59.7	62.6	50.0	48.6	75.2	84.9	85.2	96.4	61.5	76.9	64.5	96.7
GauGAN	47.0	46.5	52.1	68.5	50.9	84.3	98.6	99.8	50.1	43.6	49.5	43.8	88.8	95.9	99.4	100.0	88.5	94.4	94.2	99.8

Appendix

Table 7: Subset-level results on UnivFakeDetect. Each method reports B.Acc./AUC (%). Best and second-best results in the Average row are marked in bold and underlined, respectively.

Backbone	Supervision	$r=16$	$r=32$	$r=64$
CLIP ViT-L/14	Binary	75.42	72.20	74.89
	CuRe (Ours)	79.42	79.34	80.13
	$\Delta$	+4.00	+7.14	+5.24
DINOv2 ViT-L/14	Binary	74.32	72.52	76.57
	CuRe (Ours)	81.21	81.79	83.49
	$\Delta$	+6.89	+9.27	+6.92

Appendix

Table 8: Detailed cross-backbone results corresponding to Figure 4 . ROC-AUC scores and their differences are reported in % on AIGIBench.

\Delta

is computed by subtracting the binary supervision score from the CuRe score.

Setting	UnivFD	DRCT	DDA	GAPL	CuRe (Ours)
JPEG compression
$q=100$	74.70	73.40	85.00	91.30	92.80
$q=90$	72.19	72.00	82.00	89.14	93.09
$q=70$	73.22	71.09	77.23	85.78	92.74
$q=50$	70.64	67.73	74.17	81.87	91.72
$q=30$	64.78	64.71	70.53	74.68	90.43

Appendix

Table 9: Detailed robustness results corresponding to Figure 5 . Each entry is the 20-subset macro-average B.Acc. (%) on AIGIBench. Only one post-processing factor is varied at a time, while the other factors remain at their clean settings.

Dataset	Real/Fake	Source	#Models	Model types
AIGCDetectBenchmark ( Zhong et al., 2023 )	76.3k/76.3k	LSUN, MSCOCO, ImageNet, CelebA, FFHQ	17	GAN, diffusion
AIGIBench ( Li et al., 2025c )	84.0k/84.0k	Curated sources, social media, AI-art platforms	20	GAN, diffusion, personalized, other
UnivFakeDetect ( Ojha et al., 2023 ; Wang et al., 2020 )	45.2k/45.2k	LSUN, ImageNet, CelebA, and others	13	GAN and other
EvalGEN ( Chen et al., 2025 )	0/55.3k	Prompts	5	Diffusion, autoregressive
SynthWildX ( Cozzolino et al., 2024 )	0.5k/1.5k	X	3	Commercial text-to-image
WildRF ( Cavia et al., 2024 )	1.3k/1.3k	Reddit, Facebook, X	Unknown	Unknown

Appendix

Table 10: Test-set statistics of the ten evaluation benchmarks, including image counts, data sources, and generator coverage. Real/Fake lists the numbers of real and generated images; a combined count is marked as total. Counts above one thousand are expressed in thousands (k), rounded to one decimal place. Shared real-image pools are counted once, and additional real images used only for evaluation pairing are excluded. Unknown indicates that the generator identity or model family is unavailable for in-the-wild images.

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 198 papers

Relative Patch Response Learning for Generalizable AI-Generated Image Detection
Oct 8, 2026Tianyu Wang, Ouxiang Li, Yanbin Hao +2AI-Generated Image Detection
University of Science and Technology of China · Hefei University of Technology · University of Macau
RA-Det: Towards Universal Detection of AI-Generated Images via Robustness Asymmetry
Mar 2, 2026Xinchang Wang, Yunhao Chen, Yuechen Zhang +4AI-Generated Image Detection Adversarial Robustness
Jiangnan University · Fudan University
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

Oct 8, 2026J/K move · Enter open · S save

Relative Patch Response Learning for Generalizable AI-Generated Image Detection

Tianyu Wang, Ouxiang Li, Yanbin Hao, Zhenhua Tang+1

University of Science and Technology of China · Hefei University of Technology · University of Macau

AI-Generated Image Detection

Open paper AskPDF arXiv

Generative models can now synthesize highly realistic images, simultaneously increasing the risks of misinformation and visual forgery. Therefore, detecting AI-generated images becomes more essential, and a reliable detector must generalize to unseen generators and stay robust to unseen perturbations in the wild. Existing detectors are typically trained on either independently collected real and generated images or aligned real-generated pairs designed to mitigate content bias. Building on aligned pairs, recent methods form a mixed view by replacing some patches of the real image with their generated counterparts. However, we find that self-attention lets real and generated patches interact, so the feature of each patch no longer reflects its own source alone. This contextual shift makes a per-patch source label an imprecise target. To this end, we propose Relative Patch Response Learning (PRL). Instead of labeling each patch, PRL compares the same patch across two mixed views of an aligned pair and learns from its patch response, the change of its score between the views. (i) To give precise supervision under the contextual shift, a relative response objective measures the responses of source-changed patches against those of source-unchanged patches, which respond to the shift alone. (ii) To provide a reliable reference for the shift, a reference coherence objective keeps each group of source-unchanged patches moving as a whole. (iii) Since the two views contain different amounts of generated content, an area ranking objective asks the view with the larger generated area to have a higher mean patch score. Extensive experiments demonstrate the superior performance of PRL, which surpasses the best prior methods by 4.3% and 5.9% in average balanced accuracy across eight standard and three in-the-wild benchmarks, respectively.

Abstract

AI-Generated Image Detection

Explore similar work

Relative Patch Response Learning for Generalizable AI-Generated Image Detection

RA-Det: Towards Universal Detection of AI-Generated Images via Robustness Asymmetry

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