Relative Patch Response Learning for Generalizable AI-Generated Image Detection

Tianyu Wang; Ouxiang Li; Yanbin Hao; Zhenhua Tang; Shuo Wang

cs.CVOct 8, 2026

Relative Patch Response Learning for Generalizable AI-Generated Image Detection

Authors: Tianyu Wang, Ouxiang Li, Yanbin Hao, Zhenhua Tang, Shuo Wang

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

Abstract

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.

Figures & tables

Figure 1: Three training paradigms and the contextual shift. (a) Detectors built on hand-crafted features or on vision foundation models (VFMs) learn from real and generated images collected separately. (b) Alignment-based methods pair each real image with a reconstruction of matched content, but treat the two as independent training samples with their own image-level labels; the pair itself is not used by the supervision. (c) Patch-level methods replace some patches of a real image by their reconstruction and label each patch by its own source, in addition to the image-level label. (d) Contextual shift: a real patch $q$ has the same input in the real image and in a mixed view, yet its score rises once other patches are replaced, while the per-patch label of (c) still marks $q$ as real.

Figure 2: Contextual shift in trained detectors. Each mixed input takes the outlined square from the aligned reconstruction and the rest from the real image. The maps show how much each patch score changes from the real image to the mixed input, darker for a larger change. For DINOv3 and PPL, we train a patch head on them. Although only the square changed, the DINOv3 detector and PPL respond outside it almost as much as inside it. PRL keeps a shift outside the square and responds inside it.

Figure 3: Overview of PRL. ❶ Aligned region mixing : two masks mix the real and the synthetic image into a small and a large view. ❷ Patch responses : the patch scoring head scores every position in both views, and the patch response $\Delta_{p}$ (Eq. ( 3 )) is read on the changed sets $\mathcal{S}^{\mathrm{r,s}}$ , $\mathcal{S}^{\mathrm{s,r}}$ and on their references $\mathcal{S}^{\mathrm{r,r}}$ , $\mathcal{S}^{\mathrm{s,s}}$ ; the gating MLP gives the image-level decision. ❸ Training objectives : relative response, reference coherence and area ranking.

Method	GenImage	AIGCDetect	DRCT-2M	DDA-COCO	EvalGEN	Synthbuster	ForenSynths	UnivFD	Chameleon	SynthWildX	WildRF	Avg.
Method	Standard Benchmarks								In-the-Wild Benchmarks
UnivFD CVPR’23	64.05	72.50	59.64	52.32	15.39	55.31	77.69	78.95	50.71	52.27	55.08	57.63
FatFormer CVPR’24	61.03	74.78	50.82	51.63	27.17	53.62	86.13	82.07	51.11	53.31	59.62	59.21
NPR CVPR’24	46.74	47.78	49.62	49.07	3.43	41.83	37.88	41.92	52.26	49.80	59.34	43.61
DRCT ICML’24	80.42	67.67	90.41	66.89	50.93	73.04	54.86	64.27	68.93	74.47	74.22	69.65
C2P-CLIP AAAI’25	62.09	70.16	54.29	50.90	13.82	68.08	87.00	84.88	50.60	53.07	59.14	59.46

Table 1: Cross-benchmark comparison. Balanced accuracy (%) on eight standard and three in-the-wild benchmarks; EvalGEN has no real counterpart and is reported as accuracy on its synthetic images. Every compared method uses its official checkpoint and inference code. Bold: best per column; underlined: second best.

Method	BigGAN	CRN	CycleGAN	DeepFake	GauGAN	IMLE	ProGAN	SAN	SITD	StarGAN	StyleGAN	DALL-E	Glide-50-27	Glide-100-10	Glide-100-27	Guided	LDM-100	LDM-200	LDM-200-cfg	Avg.
UnivFD CVPR’23	87.50	55.67	96.92	69.39	98.79	68.11	99.38	58.22	62.22	95.12	79.99	79.55	77.80	77.40	77.80	67.15	91.20	90.75	67.05	78.95
FatFormer CVPR’24	97.90	69.32	99.32	86.03	99.17	69.45	99.21	62.10	80.83	99.75	87.51	88.40	70.10	69.60	70.60	61.45	88.40	89.45	70.75	82.07
NPR CVPR’24	43.25	0.00	59.83	33.30	35.62	0.00	50.45	48.40	22.22	49.47	50.18	52.30	50.50	50.85	51.00	45.75	51.15	51.10	51.05	41.92
DRCT ICML’24	52.45	49.80	48.83	56.38	49.45	49.26	53.02	86.76	54.44	54.38	54.45	79.05	60.05	66.45	63.15	59.00	94.80	94.80	94.65	64.27
C2P-CLIP AAAI’25	92.07	92.75	99.39	91.85	98.20	97.68	97.10	74.20	87.22	99.70	76.97	84.90	78.90	78.60	76.35	57.05	82.70	83.65	63.40	84.88
AIDE ICLR’25	43.68	19.96	51.73	45.54	39.73	19.38	49.31	51.37	30.56	51.43	49.57	51.80	51.25	51.55	51.50	50.25	60.80	62.25	62.35	47.05

Table 2: Comparison of balanced accuracy (%) on the UnivFD benchmark.

Table 6

Figure 4: Robustness to post-processing. Balanced accuracy on debiased GenImage-JPEG96 dataset, when every test image is JPEG-compressed at decreasing quality (a), resized by a scale factor (b) or blurred with a Gaussian of increasing

\sigma

(c).

Figure 5: Ablation. Balanced accuracy (%) on GenImage, AIGCDetect, Chameleon and WildRF. (a) Three ways of using the aligned pairs: independent image-level label without and with the mixed views, the per-patch label on the mixed views, and the patch response. (b) PRL with one objective removed.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Method	ADM	BigGAN	Glide	Midjourney	SDv1.4	SDv1.5	VQDM	Wukong	Avg.
UnivFD CVPR’23	62.51	84.37	61.28	55.06	55.57	55.67	76.88	61.10	64.05
FatFormer CVPR’24	59.82	81.77	60.28	51.99	52.81	52.76	69.97	58.86	61.03
NPR CVPR’24	46.73	44.25	46.06	43.86	46.73	47.06	50.00	49.22	46.74
DRCT ICML’24	60.74	58.38	67.86	88.68	99.37	99.21	70.03	99.10	80.42
C2P-CLIP AAAI’25	53.62	92.67	65.91	50.60	55.02	55.31	63.58	60.00	62.09
AIDE ICLR’25	50.22	50.42	52.26	57.46	75.96	76.19	50.78	69.47	60.35

Appendix

Table 5: Comparison of balanced accuracy (%) on GenImage.

Method	ADM	BigGAN	CycleGAN	DALLE2	GauGAN	Glide	Midjourney	ProGAN	SDXL	SDv1.4	SDv1.5	StarGAN	StyleGAN	StyleGAN2	VQDM	WFIR	Wukong	Avg.
UnivFD CVPR’23	62.51	87.50	96.92	49.95	98.79	61.28	55.06	99.38	58.15	55.57	55.67	95.12	79.99	69.41	76.88	69.20	61.10	72.50
FatFormer CVPR’24	59.82	97.90	99.32	49.00	99.17	60.28	51.99	99.21	63.75	52.81	52.76	99.75	87.51	81.63	69.97	87.50	58.86	74.78
NPR CVPR’24	46.73	43.25	59.83	46.65	35.62	46.06	43.86	50.45	47.48	46.73	47.06	49.47	50.18	50.02	50.00	49.65	49.22	47.78
DRCT ICML’24	60.74	52.45	48.83	68.25	49.45	67.86	88.68	53.02	80.65	99.37	99.21	54.38	54.45	53.74	70.03	50.15	99.10	67.67
C2P-CLIP AAAI’25	53.62	92.07	99.39	50.20	98.20	65.91	50.60	97.10	51.20	55.02	55.31	99.70	76.97	58.51	63.58	65.35	60.00	70.16
AIDE ICLR’25	50.22	43.68	51.73	50.95	39.73	52.26	57.46	49.31	50.55	75.96	76.19	51.43	49.57	49.20	50.78	49.20	69.47	53.98

Appendix

Table 6: Comparison of balanced accuracy (%) on AIGCDetectBenchmark.

Method	DALL-E 2	DALL-E 3	Firefly	GLIDE	Midjourney	SD1.3	SD1.4	SD2	SDXL	Avg.
UnivFD CVPR’23	70.98	34.83	77.38	40.73	39.93	57.88	57.38	63.13	55.48	55.31
FatFormer CVPR’24	53.53	35.33	59.83	67.28	43.88	52.43	53.18	51.03	66.08	53.62
NPR CVPR’24	55.76	12.36	9.46	53.01	40.81	55.76	56.41	35.86	57.11	41.83
DRCT ICML’24	47.54	53.39	48.59	60.19	89.29	94.09	94.09	90.79	79.39	73.04
C2P-CLIP AAAI’25	54.89	43.64	81.54	81.59	52.74	79.69	79.94	65.64	72.99	68.08
AIDE ICLR’25	37.83	36.63	27.73	67.93	60.48	77.03	76.68	56.18	71.38	56.87

Appendix

Table 7: Comparison of balanced accuracy (%) on Synthbuster.

Method	SDXL-Ref	SDXL-Ctrl	LCM-SDv1.5	LCM-SDXL	LDM	SD-Ctrl	SD-Turbo	SD2.1-Ctrl	SDXL-Turbo	SD2.1	SDv2-DR	SDv1-DR	SDv1.4	SDv1.5	SDXL-DR	SDXL	Avg.
UnivFD CVPR’23	54.30	72.80	53.48	65.20	81.27	66.26	55.25	63.97	52.42	56.08	52.56	56.47	55.16	54.78	52.18	62.04	59.64
FatFormer CVPR’24	48.60	61.06	48.61	49.97	52.93	49.97	48.59	50.20	48.54	48.56	53.71	54.98	48.55	48.55	51.78	48.55	50.82
NPR CVPR’24	47.72	57.20	49.16	47.72	48.85	48.11	48.28	48.57	48.07	48.50	50.78	52.60	48.84	48.77	53.03	47.76	49.62
DRCT ICML’24	85.86	80.19	99.65	78.73	99.90	99.89	91.94	95.04	70.54	96.42	92.63	99.88	99.90	99.88	72.27	83.90	90.41
C2P-CLIP AAAI’25	50.92	72.09	50.67	53.56	63.61	57.02	51.07	54.28	50.44	50.95	50.91	55.89	51.88	51.64	50.68	52.98	54.29
AIDE ICLR’25	66.05	53.42	69.95	54.13	60.74	64.80	53.03	52.95	53.07	58.09	50.02	54.52	74.82	74.28	50.01	53.11	58.94

Appendix

Table 8: Comparison of balanced accuracy (%) on DRCT-2M.

Method	DDA-COCO							EvalGEN
	VAE-ema	VAE-mse	SDXL	SD2.1	SD3.5	Flux	Avg.	Flux	GoT	Infinity	NOVA	OmniGen	Avg.
UnivFD CVPR’23	54.40	53.24	51.52	53.29	51.28	50.21	52.32	3.94	10.27	15.30	39.72	7.71	15.39
FatFormer CVPR’24	51.49	53.17	50.77	53.31	51.63	49.42	51.63	1.59	16.14	19.03	91.85	7.22	27.17
NPR CVPR’24	49.39	48.88	48.30	48.85	48.52	50.46	49.07	2.04	1.10	6.01	2.42	5.57	3.43
DRCT ICML’24	83.31	76.99	62.49	76.82	51.47	50.24	66.89	36.26	49.59	63.40	49.68	55.70	50.93
C2P-CLIP AAAI’25	51.42	51.27	50.22	51.23	51.28	49.98	50.90	1.78	6.71	11.39	46.57	2.66	13.82

Appendix

Table 9: Comparison of balanced accuracy (%) on DDA-COCO and of accuracy (%) on the synthetic images of EvalGEN.

Method	BigGAN	CRN	CycleGAN	DeepFake	GauGAN	IMLE	ProGAN	SAN	SITD	StarGAN	StyleGAN	StyleGAN2	WFIR	Avg.
UnivFD CVPR’23	87.50	55.67	96.92	69.39	98.79	68.11	99.38	58.22	62.22	95.12	79.99	69.41	69.20	77.69
FatFormer CVPR’24	97.90	69.32	99.32	86.03	99.17	69.45	99.21	62.10	80.83	99.75	87.51	81.63	87.50	86.13
NPR CVPR’24	43.25	0.00	59.83	33.30	35.62	0.00	50.45	48.40	22.22	49.47	50.18	50.02	49.65	37.88
DRCT ICML’24	52.45	49.80	48.83	56.38	49.45	49.26	53.02	86.76	54.44	54.38	54.45	53.74	50.15	54.86
C2P-CLIP AAAI’25	92.07	92.75	99.39	91.85	98.20	97.68	97.10	74.20	87.22	99.70	76.97	58.51	65.35	87.00
AIDE ICLR’25	43.68	19.96	51.73	45.54	39.73	19.38	49.31	51.37	30.56	51.43	49.57	49.20	49.20	42.36

Appendix

Table 10: Comparison of balanced accuracy (%) on ForenSynths.

Method	FLUX-Fill (rand.)	FLUX-Fill (obj.)	FLUX.2 -klein	Qwen- Edit	Step1X- Edit	Nano Banana 2	GPT- Image 2	Avg.
UnivFD CVPR’23	50.16	50.83	50.72	50.39	50.72	50.26	50.23	50.47
FatFormer CVPR’24	49.86	49.95	50.11	50.05	50.26	50.00	56.96	51.02
NPR CVPR’24	45.09	46.02	49.81	55.17	48.77	55.53	80.33	54.38
DRCT ICML’24	51.00	50.86	50.56	52.17	50.13	57.21	63.99	53.70
AIDE ICLR’25	51.13	50.65	49.63	58.63	57.54	54.43	60.43	54.63
AlignedForensics ICLR’25	50.51	50.44	50.42	73.18	50.43	64.14	56.96	56.58

Appendix

Table 11: Comparison of balanced accuracy (%) on the ManipulationBench subset of DailyBench; rows other than ours are as reported by DailyBench, including its own detector FPD.

Method	GenImage	AIGCDetect	Chameleon	WildRF	Avg.
Image-level BCE on the aligned pairs	94.58	96.06	85.89	94.06	92.65
+ mixed views, image-level BCE	97.33	97.38	82.84	93.17	92.68
+ mixed views, per-patch label	97.45	96.92	79.29	93.62	91.82
PRL (Ours)	96.83	97.75	90.01	97.59	95.55
w/o $\mathcal{L}_{\mathrm{rel}}$	96.32	96.77	83.20	95.82	93.03
w/o $\mathcal{L}_{\mathrm{coh}}$	96.05	96.43	86.83	96.89	94.05

Appendix

Table 12: Ablation. Balanced accuracy (%) of the variants of Figure 5 on GenImage, AIGCDetect, Chameleon and WildRF. All rows use the same DINOv3 backbone with LoRA and the same aligned training pairs. Top: three ways of using the aligned pairs; bottom: PRL with one objective removed.

Method	GenImage	AIGCDetect	DRCT-2M	DDA-COCO	EvalGEN	Synthbuster	ForenSynths	UnivFD	Chameleon	SynthWildX	WildRF	Avg.
Method	Standard Benchmarks								In-the-Wild Benchmarks
DDA, released model (DINOv2)	91.74	87.83	98.47	92.89	96.86	85.52	81.41	82.58	82.41	90.87	90.20	89.16
DDA on DINOv3	97.93	96.09	99.04	95.98	99.40	88.77	87.50	85.52	91.82	89.01	94.17	93.20
PRL (Ours)	96.83	97.75	99.27	94.39	99.14	97.00	92.57	94.22	90.01	93.60	97.59	95.67

Appendix

Table 13: DDA and PRL on the same backbone. Balanced accuracy (%) on the eleven benchmarks of Table 1 for the released DDA model (DINOv2), for DDA adapted from its official code to our DINOv3 backbone with LoRA, and for PRL, which uses the same backbone, adaptation and aligned pairs.

Figure 6: Sensitivity to the loss weights. Mean balanced accuracy (%) over GenImage, AIGCDetect, Chameleon and WildRF when one weight varies and the other two keep their default values; the red bar is the default setting.

Explore similar work

Aug 10, 2026cs.CV

PatchHead: Learning Spatial Patch Evidence for Generalizable AI-Generated Image Detection

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

AI-Generated Image Detection

Ask PDF

Sep 29, 2026cs.CV

Look Closer: Patch-wise Supervision for AI-Generated Image Detection

How much of an image does a detector need to see? Small RGB regions can retain useful evidence of image synthesis even when they reveal little of the full scene. Motivated by single-patch detection, we study patch-wise supervision: a shared backbone classifies explicit crops, each crop receives its own loss, and patch probabilities are averaged only at inference. The procedure requires neither handcrafted residual filtering nor a learned image-level fusion module. Experiments span single-patch selection, multiple generator collections, and four CNN and Transformer backbones. On GenImage, the reported patch-wise variants improve average accuracy over their whole-image counterparts across all four backbones. Comparisons of supervision granularity, source resolution, crop size, and inference coverage further characterize the approach, while post-processing tests and difficult-image evaluation reveal its limitations. The historical experiments include evaluation-based model selection, so their scores are not presented as a uniformly selected leaderboard comparison. Overall, the study identifies explicit local input and patch-level supervision as a simple, useful combination for investigating generalizable AI-generated image detection.

Zhida Zhang, Tao Wu, Siyu Liu +1

MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences (CASIA), Beijing, China · ShanghaiTech University, Shanghai, China · Anhui University, Hefei, China

AI-Generated Image Detection

Ask PDF

Aug 13, 2026cs.CV

SPARED: Reasoning-Based AI-Generated Image Detection via Adversarially Edited Data

Detecting AI-generated images is only half the task: a deployed detector must also justify its verdict, yet existing detectors inherit three failure modes from their training data: real and fake images collected from different sources invite provenance shortcuts, supervised explanation corpora teach templated rationales, and a static forgery corpus leaves the decision boundary standing still while generators keep moving. We introduce \methodname{}, an adversarial reinforcement learning framework that pits two heterogeneous models against each other. A diffusion image editor learns to edit real photographs into fake counterparts of those same photographs that fool the current detector, while a reasoning MLLM learns to expose them with a verdict grounded in free-form reasoning. Both rewards are shortcut-proof by design: the attacker is credited only when its edit is faithfully executed, and the defender only when its verdict is correct. As the two models alternate, each round's attacker regenerates a harder training pool aimed at the current detector's blind spots, so the detector must generalize rather than memorize any fixed artifact distribution. Although the explanation is never rewarded, its quality rises round over round as a side effect of accuracy-only training. A detector trained within this loop improves monotonically across rounds on each of three external benchmarks.

Yicheng Bao, Xiahui Guo, Xuhong Wang +1

1East China Normal University · 2Shanghai Artificial Intelligence Laboratory

Adversarial Training Adversarial Examples AI-Generated Image Detection

Ask PDF