Layer Consistency Matters: Elegant Latent Transition Discrepancy for Generalizable Synthetic Image Detection

Yawen Yang; Feng Li; Shuqi Kong; Yunfeng Diao; Xinjian Gao; Zenglin Shi; Meng Wang

cs.CVMar 11, 2026

Layer Consistency Matters: Elegant Latent Transition Discrepancy for Generalizable Synthetic Image Detection

Authors: Yawen Yang, Feng Li, Shuqi Kong, Yunfeng Diao, Xinjian Gao, Zenglin Shi, Meng Wang

Organizations: Hefei University of Technology

Abstract

Recent rapid advancement of generative models has significantly improved the fidelity and accessibility of AI-generated synthetic images. While enabling various innovative applications, the unprecedented realism of these synthetics makes them increasingly indistinguishable from authentic photographs, posing serious security risks, such as media credibility and content manipulation. Although extensive efforts have been dedicated to detecting synthetic images, most existing approaches suffer from poor generalization to unseen data due to their reliance on model-specific artifacts or low-level statistical cues. In this work, we identify a previously unexplored distinction that real images maintain consistent semantic attention and structural coherence in their latent representations, exhibiting more stable feature transitions across network layers, whereas synthetic ones present discernible distinct patterns. Therefore, we propose a novel approach termed latent transition discrepancy (LTD), which captures the inter-layer consistency differences of real and synthetic images. LTD adaptively identifies the most discriminative layers and assesses the transition discrepancies across layers. Benefiting from the proposed inter-layer discriminative modeling, our approach exceeds the base model by 14.35% in mean Acc across three datasets containing diverse GANs and DMs. Extensive experiments demonstrate that LTD outperforms recent state-of-the-art methods, achieving superior detection accuracy, generalizability, and robustness. The code is available at https://github.com/yywencs/LTD

Figures & tables

Panel of Figure 1 — Figure 1: Illustration of inter-layer feature consistency, computational efficiency, and detection performance. (a) and (b) calculate the cosine similarity and L2 distance between adjacent ViT mid-level layers, respectively. The features from real images maintain stable consistency with lower-layer transition discrepancy (LTD). (c) Inference speed measured in FPS. (d) Detection performance with recent state-of-the-art methods on the UFD [ 37 ] dataset.

Figure 2: CAM visualization of features extracted by the CLIP ViT-L/14 transformer layers for ProGAN, SD v1.4, and real images. ProGAN and SD v1.4 show noticeable attention shifts between foreground and background regions in mid-level features, indicating unstable representations and discrepancies in layer transition. Real images exhibit highly consistent attention across adjacent layers, reflecting smooth and stable evolution.

Figure 3: t-SNE maps of the layer transition discrepancy (LTD) between adjacent layers for real and generated images. We can see that 1) Shallow layers (Layer 0 vs. 1) and deep layers (Layer 22 v.s. 23) show high consistency in both real and generated images, offering limited discriminative power; 2) Middle layers (Layer 10 vs. 11 and Layer 14 vs. 15) exhibit a clearer gap of the LTD between real and fake images, providing more discriminative clues for detection.

Figure 4: Overview of the proposed Layer Transition Discrepancy(LTD) architecture. Discriminative intermediate features are extracted from a pre-trained Vision Transformer and used to construct two branches: one stacking multi-layer features directly, and the other computing pairwise differences between adjacent layers. A shared transformer block processes both branches, followed by an MLP for classification based on the concatenated representation.

Method	GAN-based Models							Deep- Fake	LDM			ADM	Glide			Dalle	Mean Acc
Method	Pro- GAN	Style- GAN	Style2- GAN	Big- GAN	Cycle- GAN	Star- GAN	Gau- GAN	Deep- Fake	100	200	200_cfg	ADM	50_27	100_10	100_27	Dalle	Mean Acc
CNNSpot [ 53 ]	100.00	85.72	83.39	70.17	85.71	91.70	78.93	53.79	55.15	53.90	53.90	61.00	62.75	61.00	60.35	56.10	69.60
UnivFD [ 37 ]	99.81	84.93	74.96	95.08	98.34	95.75	99.47	68.51	95.00	94.40	74.00	69.65	79.05	77.90	78.50	87.30	85.79
LGrad [ 52 ]	99.83	94.73	96.05	85.62	84.96	99.62	72.49	56.48	93.20	92.60	94.05	77.50	89.05	88.10	88.10	86.65	87.44
NPR [ 51 ]	99.79	97.71	98.39	84.35	96.84	99.35	82.50	80.22	98.65	98.45	98.30	75.15	97.55	97.60	97.10	93.85	93.49
FatFormer [ 31 ]	99.70	97.13	98.70	99.50	99.36	99.70	99.30	93.27	98.50	98.50	94.80	75.92	94.31	94.15	94.13	98.70	95.98

Table 1: Performance comparison on UFD in Acc. (%). Each row corresponds to a detection method, and each column to a generative model. The last column reports the mean Acc across all models. Red and blue colors indicate the best and second-best results, respectively.

Method	GAN-based Models							Deep- Fake	LDM			ADM	GLIDE			DALL-E	Mean AP
Method	Pro- GAN	Style- GAN	Style2- GAN	Big- GAN	Cycle- GAN	Star- GAN	Gau- GAN	Deep- Fake	100	200	200_cfg	ADM	50_27	100_10	100_27	DALL-E	Mean AP
CNNSpot [ 53 ]	100.00	99.54	99.06	84.51	93.43	98.15	89.49	89.02	72.50	71.11	71.11	73.96	84.6	82.01	80.47	71.27	85.01
UnivFD [ 37 ]	100.00	97.56	97.90	99.27	99.80	99.37	99.98	81.79	99.27	99.32	92.50	87.64	95.56	94.96	95.28	97.47	96.10
LGrad [ 52 ]	100.00	99.86	99.89	90.73	93.61	99.98	79.28	72.67	98.36	98.18	98.57	87.06	94.53	94.20	93.23	95.72	93.49
NPR [ 51 ]	99.99	99.78	99.94	87.80	98.94	99.94	85.49	82.40	99.80	99.76	99.73	80.29	99.61	99.63	99.55	98.96	95.73
FatFormer	100.00	99.75	99.92	99.98	99.99	100.00	100.00	97.99	99.89	99.83	99.22	91.92	99.50	99.33	99.27	99.84	99.15

Table 2: Performance comparison on UFD in AP. Each row corresponds to a detection method, and each column to a generative model. The last column reports the mean AP across all models. Red and blue colors indicate the best and second-best results, respectively.

Method	SD variants						Turbo Variants		LCM Variants		ControlNet Variants			Mean Acc
Method	LDM	SDv1.4	SDv1.5	SDv2	SDXL	SDXL- Refiner	SD- Turbo	SDXL- Tubo	LCM- SDv1.5	LCM- SDXL	SDv1- Ctrl	SDv2- Ctrl	SDXL- Ctrl	Mean Acc
CNNSpot [ 53 ]	74.26	99.87	99.88	87.44	54.04	54.48	84.00	80.04	98.59	64.10	79.46	63.11	68.21	77.50
UnivFD [ 37 ]	97.93	95.40	93.73	93.73	89.25	91.47	86.20	84.73	94.64	89.87	88.71	80.86	74.89	89.34
LGrad [ 52 ]	52.07	51.46	51.49	50.38	50.16	50.27	50.07	50.91	50.24	49.99	50.10	50.47	51.77	50.72
NPR [ 51 ]	50.31	99.70	99.65	54.51	50.01	50.17	99.30	99.61	99.92	50.17	61.47	53.99	64.76	71.81
DIRE [ 54 ]	57.43	98.95	98.85	65.87	59.63	63.95	89.48	88.00	97.42	70.35	90.28	85.67	82.67	80.66

Table 3: Performance comparison on DRCT-2M in Acc. (%). Each row corresponds to a detection method, and each column to a generative model. The last column reports the mean Acc across all models. Red and blue colors indicate the best and second-best results, respectively.

Method	SD variants						Turbo Variants		LCM Variants		ControlNet Variants			Mean AP
Method	LDM	SDv1.4	SDv1.5	SDv2	SDXL	SDXL- Refiner	SD- Turbo	SDXL- Tubo	LCM- SDv1.5	LCM- SDXL	SDv1- Ctrl	SDv2- Ctrl	SDXL- Ctrl	Mean AP
CNNSpot [ 53 ]	91.53	100.00	100.00	99.68	92.87	92.34	99.47	98.94	99.97	87.39	92.37	97.12	86.63	95.25
UnivFD [ 37 ]	99.89	99.72	99.72	99.62	99.28	99.55	98.96	98.80	99.67	99.26	99.24	98.25	97.53	99.19
LGrad [ 52 ]	52.63	60.23	59.36	57.18	44.30	51.36	50.30	46.50	49.86	39.04	49.61	51.94	54.58	51.30
NPR [ 51 ]	99.60	99.98	99.98	97.17	70.99	76.71	100.00	100.00	100.00	84.86	99.05	97.81	99.29	94.26
DIRE [ 54 ]	90.43	99.96	99.96	93.13	92.63	95.16	99.45	99.28	99.89	96.68	99.33	99.21	98.97	97.24

Table 4: Performance comparison on DRCT-2M in AP. Each row corresponds to a detection method, and each column to a generative model. The last column reports the mean AP across all models. Red and blue colors indicate the best and second-best results, respectively.

Method	ADM	Midjourney	GLID	VQDM	SDv1.4	SDv1.5	Wukong	BigGAN	Mean Acc
CNNSpot [ 53 ]	58.77	52.58	55.00	53.67	51.58	51.83	50.25	79.54	56.65
UnivFD [ 37 ]	66.87	56.13	62.46	85.31	63.66	63.49	70.93	90.13	69.87
LGrad [ 52 ]	67.12	58.98	62.98	70.08	68.41	69.00	65.48	81.32	67.92
NPR [ 51 ]	69.72	78.00	78.35	78.13	78.62	78.88	76.10	80.30	77.26
FatFormer [ 31 ]	79.49	55.52	89.09	87.98	88.09	87.88	88.03	98.67	84.34
VIB [ 60 ]	72.20	58.50	69.75	87.15	67.50	67.50	74.00	88.55	73.14

Table 5: Performance comparison on GenImage in Acc. (%). Each row corresponds to a detection method, and each column to a generative model. The last column reports the mean Acc across all models. Red and blue colors indicate the best and second-best results, respectively.

Method	ADM	Midjourney	GLID	VQDM	SDv1.4	SDv1.5	Wukong	BigGAN	Mean AP
CNNSpot [ 53 ]	71.10	55.94	66.19	61.96	56.91	51.83	50.25	89.89	63.01
UnivFD [ 37 ]	86.81	74.00	83.81	96.53	86.14	85.84	91.07	98.13	87.79
LGrad [ 52 ]	72.64	69.13	76.28	74.36	71.85	72.51	68.03	88.83	74.20
NPR [ 51 ]	74.64	85.60	85.72	81.79	84.03	84.59	80.47	88.40	83.16
FatFormer [ 31 ]	95.00	73.78	97.72	98.53	98.37	98.14	98.62	99.93	95.01
VIB [ 60 ]	86.70	74.41	88.66	96.22	85.13	83.99	90.15	97.29	87.82

Table 6: Performance comparison on GenImage in AP. Each row corresponds to a detection method, and each column to a generative model. The last column reports the mean AP across all models. Red and blue colors indicate the best and second-best results, respectively.

Figure 5: Robust evaluation of our method and baselines under (a) JPEG compression with different quality factors and (b) downsampling with different downscaling factors. Our LTD validates superior robustness to different degradation types and levels.

Raw mid-level	LTD	Pos. Enc.	UFD	DRCT-2M	Mean Acc.
✓	✗	✗	84.92	92.75	88.84
✓	✗	✓	94.22	96.12	95.17
✗	✓	✗	86.42	93.50	89.96
✗	✓	✓	92.43	94.01	93.22
✓	✓	✓	96.90	99.54	98.22

Table 7: Ablation study on the layer transition consistency (LTD) detection. Raw ML. denotes raw mid-level features and Pos. Enc. denotes positional encoding, respectively.

Layer Selection	UFD	DRCT-2M	GenImage	Mean Acc
Fixed(0-7)	62.87	82.12	45.70	63.56
Fixed(8-15)	95.78	98.98	90.38	95.50
Fixed(16-24)	92.62	97.57	79.78	89.99
Fixed(7-19)	94.89	88.67	90.72	91.43
Fixed(9-19)	95.02	95.75	91.34	94.04
Fixed(11-19)	96.26	97.44	91.63	95.11

Table 8: Ablation study on the number of selected layers. Mean Acc is reported across different datasets.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Training	CNNSpot	LGrad	UnivFD	NPR	AIDE	D 3	ForgeLens	LTD
Dataset	[ 53 ]	[ 52 ]	[ 37 ]	[ 51 ]	[ 56 ]	[ 12 ]	[ 6 ]	(ours)
ProGAN	57.31	59.41	57.36	59.83	58.38	56.63	57.71	58.27
ProGAN	99.71/0.89	99.44/6.13	97.27/4.25	99.56/5.47	98.47/5.04	99.06/0.18	77.33/31.46	98.55/4.66
SDXL-Turbo	57.09	55.76	63.78	58.37	56.78	72.96	49.22	73.98
SDXL-Turbo	100.00/0.00	95.08/3.45	44.45/83.11	95.69/8.71	98.71/0.98	63.30/85.81	5.55/92.64	67.63/82.43
LCM-SDv1.5	61.96	53.30	68.00	60.28	62.60	70.50	57.09	75.66

Appendix

Table 1: Accuracy (Acc.) of different detectors (columns) on the Chameleon dataset when trained with different sources (rows). For each training dataset, the first row reports the overall classification accuracy, while the second row presents the class-wise accuracy split into “fake/real” for a more detailed breakdown.

Figure 1: t-SNE visualization of feature separability under different representations. Comparison among (a) the ViT last-layer features, (b) the concatenated selected mid-layer features, and (c) our LTD-enhanced representations. While both (a) and (b) exhibit limited separation between real and generated images, (c) shows clear and robust class separation, demonstrating the effectiveness of our LTD.

Figure 2: t-SNE visualization of feature distributions on SD v1.4 generated images. Columns show features for original images, (b) JPEG-compressed images, and downsampled iamges. Under JPEG compression, ForgeLens suffers from severe cluster collapse. Notably, under downsampling, while ForgeLens retains visual separability, it exhibits a significant distribution shift that is mismatched with with the classifier (resulting in 50% Acc).

Panel of Figure 3 — Figure 3: Backbone comparison across diverse generative models. Detection accuracy of four representative backbones, including CLIP ViT-L/14, CLIP ViT-B/16, ImageNet ViT-L/16, and Flava, evaluated over a wide range of GAN- and diffusion-based generators. The results highlight significant performance differences driven by pre-training strategy and model capacity.

JPEG	Downsampling	Blur	Acc.(%)	AP (%)
✗	✗	✗	75.66	78.00
✓	✗	✗	74.46 (1.20 $\downarrow$ )	77.16 (0.84 $\downarrow$ )
✗	✓	✗	75.48 (0.18 $\downarrow$ )	77.41 (0.59 $\downarrow$ )
✗	✗	✓	69.14 (6.52 $\downarrow$ )	71.27 (6.73 $\downarrow$ )
✓	✓	✓	70.89 (4.77 $\downarrow$ )	70.63 (7.37 $\downarrow$ )

Appendix

Table 2: Robustness analysis against image degradations on the Chameleon dataset. We evaluate the performance using average accuracy (Acc.) and average precision (AP). The values in parentheses denote the performance drop compared to the clean baseline.

Layer Selection	UFD	DRCT-2M	GenImage	Mean Acc
Fixed(11-23)	95.99	91.26	91.70	92.98
Fixed(11-21)	94.47	87.09	94.55	92.04
Fixed(11-19)	95.11	91.63	97.44	94.73
Fixed(11-17)	97.03	91.11	94.72	94.29

Appendix

Table 3: Ablation study on the selection endpoint (upper bound).

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

Jul 12, 2026

Effective Synthetic Image Detection via Noise Residual Clustering

The rapid advancement of generative artificial intelligence (AI) has made synthetic images remarkably realistic, posing security threats such as misinformation and fraud. It is significant to detect the synthetic image in the manner of passive and blind image authentication. Most existing detectors rely on supervised training with large labeled datasets, leading to high costs and degraded performance on unknown generative models. To attenuate such deficiencies, we propose a training-free detection method. Specifically, noise residual fingerprints are first extracted by a simple yet effective pre-trained Noiseprint++ model. Then multi-scale features are further extracted from such residual by a frozen Vision Transformer (ViT), followed by adaptive weighted fusion. Only a few real image samples are used needed to initialize the clustering centers for unsupervised K-Means, distinguishing real and synthetic images without training. Extensive evaluations on four benchmark datasets show that our proposed scheme achieves an average accuracy of 82.2%, outperforming the state-of-the-art detectors on generalization ability. Superior performance is gained on the popular diffusion type of synthetic images, and the effectiveness of each module is validated by ablation studies. Source code will be publicly available at https://github.com/multimediaFor/NoiseCluSID.

Caihui Yan, Gang Cao, Huawei Tian +2

School of Computer and Cyber Sciences, Communication University of China, Beijing 100024, China · People's Public Security University of China, Beijing 100038, China

AI-Generated Image Detection

Ask PDF

Apr 30, 2026

HiMix: Hierarchical Artifact-aware Mixup for Generalized Synthetic Image Detection

The rapid evolution of generative models has enabled the creation of highly realistic and diverse synthetic images, posing significant challenges to reliable and generalizable Synthetic Image Detection (SID). However, existing detectors are typically trained on limited and biased datasets, resulting in poor generalization to unseen generators. To address this issue, we propose HiMix, a unified framework that enhances generalization by expanding the training distribution and promoting artifact-aware representations. Specifically, the Mixup-driven Distributional Augmentation (MDA) module constructs continuous transitional samples between real and fake images, improving coverage of low-confidence regions and exposing the model to more challenging samples, while the pixel-wise mixup operation smoothly perturbs semantics to enhance sensitivity to low-level artifacts. Moreover, the Hierarchical Artifact-aware Representation (HAR) module aggregates artifact information from both global and local levels through cross-layer integration and coarse-to-fine feature fusion, enabling the extraction of discriminative forgery representations under diverse distributions. Extensive experiments across multiple benchmarks demonstrate that HiMix achieves state-of-the-art performance, establishing well-separated logits for improved generalization to unseen forgeries.

Shuchang Zhou, Kaiwen Shen, Jiwei Wei +3

University of Electronic Science and Technology of China · Hainan University

Synthetic Data Augmentation Image Forgery Detection Data Augmentation AI-Generated Image Detection Mixup

Ask PDF

Jul 7, 2026

Generalized Synthetic Image Detection with Enhanced RGB-Noise Representation Learning

The rapid advancement of large-scale generative models has accelerated the spread of highly deceptive AI-generated images, making generalized synthetic image detection a critical imperative. Existing forensic networks often struggle with cross-model generalization and realworld degradations due to their reliance on single-domain representations and conventional binary classification optimization. To overcome these limitations, we propose RNSIDNet, a novel forensic framework that achieves robust detection through enhanced RGB-Noise representation learning. Specifically, our method employs a dual-branch architecture where global RGB semantics, extracted by an attention-refined CLIP backbone, dynamically modulate highfrequency noise artifacts captured by Bayar convolutions via a Feature-wise Linear Modulation (FiLM) module. To further enhance the learned representations, we design a Hard Sample-aware Contrastive Learning (HSCL) strategy. By explicitly penalizing challenging training samples, HSCL reshapes the latent feature space to maximize the discriminative margin between pristine and synthetic domains. Extensive experiments across eight public benchmark datasets verify that our model achieves state-of-the-art performance, delivering superior generalization ability, robustness, and computational efficiency. Code and dataset will be publicly available on https://github.com/multimediaFor/RNSIDNet.

Zhen Li, Gang Cao, Tian Zhang +2

School of Computer and Cyber Sciences, Communication University of China, Beijing, 100024, China · School of Information Engineering, Changsha Medical University, Changsha, 410219, China · Department of Information Engineering, Beijing Institute of Graphic Communication, Beijing, 100026, China +1

Contrastive Learning Image Forgery Detection AI-Generated Image Detection

Ask PDF

Abstract

AI-Generated Image Detection

Explore similar work

Effective Synthetic Image Detection via Noise Residual Clustering

HiMix: Hierarchical Artifact-aware Mixup for Generalized Synthetic Image Detection

Generalized Synthetic Image Detection with Enhanced RGB-Noise Representation Learning