Improving Anomaly Detection with Foundation-Model Synthesis and Wavelet-Domain Attention

Wensheng Wu; Zheming Lu; Ziqian Lu; Zewei He; Xuecheng Sun; Zhao Wang; Jungong Han; Yunlong Yu

cs.CVMar 3, 2026

Improving Anomaly Detection with Foundation-Model Synthesis and Wavelet-Domain Attention

Authors: Wensheng Wu, Zheming Lu, Ziqian Lu, Zewei He, Xuecheng Sun, Zhao Wang, Jungong Han, Yunlong Yu

Organizations: School of Aeronautics and Astronautics, Zhejiang University, Hangzhou, 310027, China · School of Computer Science and Technology(), , Hangzhou, 310018, China · Ningbo Innovation Center, Zhejiang University, Ningbo, China · Department of Automation, Tsinghua University, Beijing, 100084, China ·

Abstract

Industrial anomaly detection faces significant challenges due to the scarcity of anomalous samples and the complexity of real-world anomalies. In this paper, we propose a foundation model-based anomaly synthesis pipeline (FMAS) that generates highly realistic anomalous samples without fine-tuning or class-specific training. Motivated by the distinct frequency-domain characteristics of anomalies, we introduce aWavelet Domain Attention Module (WDAM), which exploits adaptive sub-band processing to enhance anomaly feature extraction. The combination of FMAS and WDAM significantly improves anomaly detection sensitivity while maintaining computational efficiency. Comprehensive experiments on MVTec AD and VisA datasets demonstrate that WDAM, as a plug-and-play module, achieves substantial performance gains against existing baselines.

Figures & tables

Figure 1: Qualitative results showing anomaly samples generated by the proposed FMAS on the MVTec AD dataset, with red contours highlighting the anomalous regions.

Figure 2: Visualization of anomaly saliency in the four wavelet sub-bands. The leftmost column presents the original images, the middle columns show the LL, LH, HL, and HH sub-bands, and the rightmost column illustrates the corresponding masks. Red dashed boxes indicate defect locations, and red solid boxes highlight magnified regions. All samples are drawn from the MVTec AD dataset.

Class	bottle	cable	capsule	carpet	grid	hazelnut	leather	metal nut	pill	screw	tile	toothbrush	transistor	wood	zipper
LL	0.69	0.84	0.90	0.72	0.76	0.76	0.69	0.78	0.81	0.75	0.83	0.80	0.81	0.70	0.70
LH	0.10	0.07	0.05	0.12	0.10	0.08	0.14	0.10	0.08	0.10	0.08	0.08	0.07	0.09	0.12
HL	0.10	0.07	0.04	0.13	0.10	0.08	0.13	0.09	0.08	0.10	0.08	0.10	0.08	0.17	0.14
HH	0.10	0.02	0.02	0.03	0.03	0.08	0.04	0.03	0.03	0.04	0.02	0.03	0.04	0.04	0.04

Table 1: Quantitative comparison of anomaly prominence across different wavelet sub-bands on our synthesized dataset based on the MVTec AD dataset. The reported values represent the relative degree of pixel-level change within each sub-band after anomaly synthesis, with higher values indicating greater relative variation.

Figure 3: Illustration of the proposed foundation model-based anomaly synthesis pipeline (FMAS).

Figure 4: The proposed Wavelet Domain Attention Module (WDAM) processes input feature maps through a frequency-aware attention mechanism. First, WDAM decomposes the spatial features into multiple frequency sub-bands using Discrete Wavelet Transform (DWT). An adaptive attention mechanism then selectively enhances or suppresses features in each sub-band based on their relevance to anomaly detection. Finally, the refined frequency components are reconstructed back into the spatial domain via Inverse Discrete Wavelet Transform (IDWT), effectively preserving critical defect-related patterns while suppressing noise.

Metric	GLASS ( Chen et al., 2024 )	RealNet ( Zhang et al., 2024 )	Ours
FID $\downarrow$	177.51	69.64	56.99
LPIPS $\downarrow$	0.280	0.147	0.130

Table 2: Quantitative results of anomaly image realism between our method and the alternative anomaly synthesis approaches on the MVTec-AD dataset.

Figure 5: (a) The architecture of CutPaste with the proposed WDAM, wherein Bottleneck* denotes a modified bottleneck block incorporating WDAM, as elaborated in (b). The elements enclosed within the red dashed box indicate the additional components introduced in comparison to the original bottleneck configuration.

Category	CutPaste	+FMAS	+FM	+WDAM (Final Model)
Bottle	( 100.0 , 68.06, 39.84)	( 100.0 , 81.18, 66.07)	( 100.0 , 83.73 , 71.35 )	( 100.0 , 86.82 , 71.29 )
Cable	(88.44, 80.39, 43.13)	( 94.10 , 87.80 , 59.77)	(91.55, 85.48, 62.97 )	( 97.91 , 90.79 , 74.14 )
Capsule	( 100.0 , 89.67, 65.91)	(96.38, 92.58 , 74.94)	(96.66, 91.04, 75.15 )	( 97.50 , 93.26 , 77.39 )
Carpet	(81.77, 88.02, 23.36)	( 92.12 , 95.97 , 68.20 )	(91.59, 95.59, 62.93)	( 98.66 , 96.70 , 81.25 )
Grid	(90.64, 77.65, 56.42)	(99.14, 92.08 , 68.69 )	( 99.16 , 90.69 , 63.96)	( 99.89 , 88.78, 66.29 )
Hazelnut	(98.58, 84.77, 69.51)	( 99.48 , 92.31 , 82.12 )	(98.89, 91.40, 73.78)	( 99.73 , 92.54 , 84.02 )

Table 3: Evaluation results of CutPaste and our variants on MVTec AD dataset, with image-level AUROC, pixel-level AUROC, and PRO metrics. FM is short for foreground mask. Bold and underlined indicate the best and second-best results, respectively.

Category	CutPaste	+FMAS	+FM	+WDAM (Final Model)	EawT ( Du et al., 2024 )	AST ( Rudolph et al., 2023 )
candle	(94.7, 93.5)	(93.1, 94.8)	(94.3, 95.8 )	( 98.0 , 93.4)	(86.8, 98.6 )	( 97.1 , 93.6)
capsules	(84.4, 83.9 )	(72.7, 74.7)	(80.7, 79.0)	( 85.2 , 77.7)	( 91.2 , 99.3 )	(81.5, 80.1)
cashew	(94.9, 88.3)	(90.2, 90.5)	( 95.3 , 96.0 )	( 95.6 , 92.7 )	(88.4, 81.8)	(93.0, 86.3)
chewinggum	(97.7, 97.9 )	(97.5, 97.2)	( 99.1 , 93.3)	(99.0, 85.3)	(98, 98.8 )	( 99.3 , 87.7)
fryum	(93.9, 97.0 )	(90.4, 88.1)	(95.8, 93.9)	( 97.1 , 95.4 )	( 97.5 , 95.4 )	(96.8, 79.3)
macaroni1	(85.1, 78.2)	(81.9, 90.2)	( 97.0 , 91.5 )	( 95.6 , 91.0)	(89.4, 99.9 )	(75.0, 89.7)

Table 4: Evaluation results of CutPaste, its variants with different modifications, and the other competitors on VisA dataset, with image-level and pixel-level AUROC metrics

Category	CFLOW-AD ( Gudovskiy et al., 2022 )	MSTUnet ( Jiang et al., 2022 )	MambaAD ( He et al., 2024 )	DRAEM ( Zavrtanik et al., 2021 )	+SSPCAB ( Ristea et al., 2022 )	Final Model
Bottle	(99.0, 96.8)	( 100.0 , 99.0 )	( 100.0 , 98.8 )	( 99.5 , 98.5)	(98.4, 98.8 )	(99.2, 98.4)
Cable	( 97.6 , 93.5)	(91.4, 89.9)	( 98.8 , 95.8)	(95.3, 94.4)	(96.9, 96.0 )	(97.4, 96.5 )
Capsule	( 99.0 , 93.4)	(98.4, 95.7 )	(94.4, 98.4 )	(96.6, 90.8)	( 99.3 , 93.1)	(97.0, 91.6)
Carpet	(99.3, 97.7)	( 99.9 , 98.3 )	( 99.8 , 99.2 )	(97.9, 95.3)	(98.2, 95.0)	(99.6, 95.8)
Grid	( 99.0 , 96.1)	( 100.0 , 99.7 )	( 100.0 , 99.2)	( 100.0 , 99.7 )	( 100.0 , 99.5 )	( 100.0 , 99.7 )
Hazelnut	( 99.0 , 96.7)	( 100.0 , 99.3)	( 100.0 , 99.0)	( 100.0 , 99.7 )	( 100.0 , 99.8 )	( 100.0 , 99.6)

Table 5: Image-level AUROC and pixel-Level AUROC results on the MVTec AD Dataset. Comparison of DRAEM ( Zavrtanik et al., 2021 ) , DRAEM with the proposed WDAM, and other methods, including CFLOW-AD ( Gudovskiy et al., 2022 ) , MSTUnet ( Jiang et al., 2022 ) , MambaAD ( He et al., 2024 ) and DRAEM+SSPCAB ( Ristea et al., 2022 )

Dataset	THFR ( Guo et al., 2023 )	RD++ ( Tien et al., 2023 )	Dinomaly ViT-B ( Guo et al., 2025 )	PatchCore	PatchCore Ensemble	Final Model
MVTec AD	(99.2, 98.2, 95.0)	(99.4, 98.3 , 95.0)	( 99.6 , 98.4 , 94.8)	(99.1, 98.1, 97.4)	( 99.5 , 98.2, 97.6 )	(99.0, 98.3 , 97.7 )

Table 6: Evaluation results of PatchCore, its variants with different modifications, and other methods, including THFR ( Guo et al., 2023 ) , RD++ ( Tien et al., 2023 ) , and Dinomaly ViT-B ( Guo et al., 2025 ) , on the MVTec AD dataset using image-level AUROC, pixel-level AUROC, and PRO metrics.

Figure 6: Qualitative results of CutPaste ( Li et al., 2021 ) and our method on MVTec AD dataset. From top to down, original image, ground-truth mask, detection heatmap of CutPaste, and detection heatmap of our method.

Figure 7: Qualitative results of CutPaste ( Li et al., 2021 ) and our method on VisA dataset. From top to down, original image, ground-truth mask, detection heatmap of CutPaste, and detection heatmap of our method.

Dataset	FMAS	WDAM	Metrics
MVTec AD			(93.23, 82.34, 57.51)
	✓		( 96.54 , 86.88 , 67.69 )
		✓	(96.20, 84.91, 66.35)
	✓	✓	( 98.00 , 88.28 , 72.04 )
VisA			(90.68, 89.68, 65.65)
	✓		( 93.86 , 93.37 , 75.82 )

Table 7: Module-level ablation results on MVTec AD and VisA dataset, based on the CutPaste.

Layer1	Layer2	Layer3	Layer4	Metrics
				(96.54, 86.88, 67.69)
✓				( 98.00 , 88.28 , 72.04 )
✓	✓			( 97.92 , 87.99 , 71.70)
✓	✓	✓		(97.39, 87.92, 72.09 )
✓	✓	✓	✓	(97.45, 87.70, 71.60)

Table 8: Ablation analysis of WDAM integration within CutPaste on the MVTec AD dataset.

Metrics	WideResNet50-2	+WDAM
Param. (MB)	67.40	67.59
FLOPs (G)	29.92	29.94
Inference Time (ms)	5.71	7.35
I-AUROC (%)	96.54	98.00

Table 9: Comparison of WideResNet50-2 and WideResNet50-2+WDAM in terms of model complexity, efficiency, and accuracy.

Class	bottle	cable	capsule	carpet	grid	hazelnut	leather	metal nut	pill	screw	tile	toothbrush	transistor	wood	zipper
LL	0,0,1	0,1,0	0,1,1	0,1,0	0,1,1	0,1,0	0,1,0	0,1,1	0,0,1	0,1,0	0,1,0	0,1,0	1,1,1	0,1,1	0,1,0
LH	1,1,1	1,0,0	1,0,1	0,0,0	1,0,0	1,0,1	1,0,1	1,0,0	0,0,1	0,0,0	1,0,0	1,0,1	1,0,1	1,0,0	1,0,0
HL	1,1,0	0,0,0	1,1,0	0,1,0	0,1,0	0,0,0	1,1,0	0,1,0	0,1,0	0,0,0	0,0,1	0,0,0	1,1,1	0,0,0	0,0,1
HH	0,0,0	0,0,0	1,0,0	0,0,1	1,0,0	0,0,1	0,0,0.8	0,0,1	0,0,0	0,1,0	1,0,1	1,0,1	0,0,0	0,0,0	1,0,0

Table 10: Average wavelet sub-bands weights learned by the WDAM on the MVTec AD dataset. Each sample is assigned three groups of weights (LL, LH, HL, HH), representing the relative contributions of the four wavelet sub-bands to anomaly detection. For each class, the three comma-separated values correspond to the three WDAM modules integrated sequentially into three bottleneck layers, as illustrated in Figure 5 . Values are averaged across all samples within each module.

Industrial Anomaly Detection

Industrial anomaly detection finds defective products and abnormal equipment behavior from visual, sensor, audio, and other production data.

See all 261 papers

One-to-More: High-Fidelity Training-Free Anomaly Generation with Attention Control
Mar 18, 2026Haoxiang Rao, Zhao Wang, Chenyang Si +4Synthetic Anomaly Generation Industrial Anomaly Detection
Nanjing University · China Mobile Zijin Innovation Institute · Shandong University of Science and Technology
ChronosAD: Leveraging Time Series Foundation Models for Accurate Anomaly Detection
May 31, 2026Uzair Khan, Luigi Capogrosso, Francesco Biondani +4Industrial Anomaly Detection Time-Series Anomaly Detection
University of Verona · Interdisciplinary Transformation University of Austria · ETH Zurich
CORE: Contrast-and-Reflect for Self-Supervised Prompt Optimization in Anomaly Synthesis
Oct 14, 2026Haoxiang Rao, Fang Zhao, Zhao Wang +3Synthetic Anomaly Generation Synthetic Data Augmentation
School of Intelligence Science and Technology, Nanjing University · China Mobile Zijin Innovation Institute

Mar 18, 2026J/K move · Enter open · S save

One-to-More: High-Fidelity Training-Free Anomaly Generation with Attention Control

Haoxiang Rao, Zhao Wang, Chenyang Si, Yan Lyu+3

Nanjing University · China Mobile Zijin Innovation Institute · Shandong University of Science and Technology

Synthetic Anomaly Generation Industrial Anomaly Detection Diffusion-Based Image Generation Attention Control in Diffusion Models Few-Shot Anomaly Detection

Open paper AskPDF arXiv

Industrial anomaly detection (AD) is characterized by an abundance of normal images but a scarcity of anomalous ones. Although numerous few-shot anomaly synthesis methods have been proposed to augment anomalous data for downstream AD tasks, most existing approaches require time-consuming training and struggle to learn distributions that are faithful to real anomalies, thereby restricting the efficacy of AD models trained on such data. To address these limitations, we propose a training-free few-shot anomaly generation method, namely O2MAG, which leverages the self-attention in One reference anomalous image to synthesize More realistic anomalies, supporting effective downstream anomaly detection. Specifically, O2MAG manipulates three parallel diffusion processes via self-attention grafting and incorporates the anomaly mask to mitigate foreground-background query confusion, synthesizing text-guided anomalies that closely adhere to real anomalous distributions. To bridge the semantic gap between the encoded anomaly text prompts and the true anomaly semantics, Anomaly-Guided Optimization is further introduced to align the synthesis process with the target anomalous distribution, steering the generation toward realistic and text-consistent anomalies. Moreover, to mitigate faint anomaly synthesis inside anomaly masks, Dual-Attention Enhancement is adopted during generation to reinforce both self- and cross-attention on masked regions. Extensive experiments validate the effectiveness of O2MAG, demonstrating its superior performance over prior state-of-the-art methods on downstream AD tasks.

Abstract

Industrial Anomaly Detection

Explore similar work

One-to-More: High-Fidelity Training-Free Anomaly Generation with Attention Control

ChronosAD: Leveraging Time Series Foundation Models for Accurate Anomaly Detection

CORE: Contrast-and-Reflect for Self-Supervised Prompt Optimization in Anomaly Synthesis