SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling

Camile Lendering; Erkut Akdag; Egor Bondarev

cs.CVFeb 26, 2026

SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling

Authors: Camile Lendering, Erkut Akdag, Egor Bondarev

Organizations: AIMS Group, Department of Electrical Engineering, Eindhoven University of Technology

Abstract

Detecting visual anomalies in industrial inspection often requires training with only a few normal images per category. Recent few-shot methods achieve strong results employing foundation-model features, but typically rely on memory banks, auxiliary datasets, or multi-modal tuning of vision-language models. We therefore question whether such complexity is necessary given the feature representations of vision foundation models. To answer this question, we introduce SubspaceAD, a training-free method, that operates in two simple stages. First, patch-level features are extracted from a small set of normal images by a frozen DINOv2 backbone. Second, a Principal Component Analysis (PCA) model is fit to these features to estimate the low-dimensional subspace of normal variations. At inference, anomalies are detected via the reconstruction residual with respect to this subspace, producing interpretable and statistically grounded anomaly scores. Despite its simplicity, SubspaceAD achieves state-of-the-art performance across one-shot and few-shot settings without training, prompt tuning, or memory banks. In the one-shot anomaly detection setting, SubspaceAD achieves image-level and pixel-level AUROC of 97.1% and 97.5% on the MVTec-AD dataset, and 93.2% and 98.2% on the VisA dataset, respectively, surpassing prior state-of-the-art results. Code and demo are available at https://github.com/CLendering/SubspaceAD.

Figures & tables

Figure 1: One-shot segmentation results of SubspaceAD on the MVTec-AD dataset [ 3 ] , where SubspaceAD only uses one normal image per category. Each example shows a test sample with its predicted anomaly mask (overlaid in dark blue), across all 15 categories of the MVTec-AD dataset.

Figure 2: Overview of SubspaceAD. (Fitting) : Aggregated patch features are collected from $k$ normal samples using a frozen DINOv2-G model and a PCA model is fitted to capture the subspace of normal variation. (Inference) : Features of a test image are extracted, projected onto the normal subspace, and the reconstruction error is computed, providing the anomaly segmentation map directly. PCA figure from [ 38 ] .

		MVTec-AD				VisA
		Image-level		Pixel-level		Image-level		Pixel-level
Setup	Method	AUROC	AUPR	AUROC	PRO	AUROC	AUPR	AUROC	PRO
0-shot	WinCLIP	91.8	96.5	85.1	64.6	78.1	81.2	79.6	56.8
1-shot	SPADE	81.0	90.6	91.2	83.9	79.5	82.0	95.6	84.1
	PatchCore	83.4	92.2	92.0	79.7	79.9	82.8	95.4	80.5
	FastRecon	-	-	-	-	-	-	-	-

Table 1: Comparison of anomaly detection and localization performance on MVTec-AD and VisA across different few-shot settings. Results for SubspaceAD (ours) are reported as mean

\pm

standard deviation over 5 seeds. Best results are in bold , and second-best are underlined .

Panel of Figure 3 — Figure 3: Qualitative comparison on VisA and MVTec-AD (1-shot). SubspaceAD produces sharper and more precise anomaly maps than PromptAD [ 22 ] and AnomalyDINO [ 11 ] , with fewer false activations and better alignment with ground-truth defects across both datasets. More qualitative examples are provided in Appendix E .

Zero-shot methods
Method	MVTec-AD	VisA
WinCLIP [ 18 ]	91.8	78.1
AnomalyCLIP [ 42 ]	91.5	82.1
Batched zero-shot methods
ACR [ 20 ]	85.8	–
MuSc [ 23 ]	97.8	94.1

Table 2: Batched 0-shot anomaly detection. All values are image-level AUROC (%). Best results are in bold , second-best are underlined .

Figure 4: Effect of image resolution on performance across both datasets. Performance peaks at 672 px on both MVTec-AD (solid) and VisA (dashed).

	MVTec-AD		VisA
Strategy	I-AUROC	PRO	I-AUROC	PRO
Mean-pool (Middle-7)	98.4	95.0	94.1	95.1
Mean-pool (Final-7)	98.2	93.3	91.9	93.5
Concat (Middle-7)	98.6	95.0	93.8	95.0
Last layer only	97.6	92.5	89.3	91.2

Table 3: Evaluation of feature aggregation strategies (4-shot). The selected approach Mean-pool (Middle-7) achieves the best overall performance, with image-level AUROC and PRO given in (%).

Figure 5: Impact of backbone scale on SubspaceAD performance on both datasets. Performance improves with increasing model capacity, indicating that richer foundation features directly enhance few-shot anomaly detection.

I-AUROC (%)
	MVTec-AD			VisA
$\tau$	$k{=}1$	$k{=}2$	$k{=}4$	$k{=}1$	$k{=}2$	$k{=}4$
0.95	98.0	98.2	98.4	92.2	93.4	93.7
0.96	98.1	98.2	98.4	92.4	93.5	93.9
0.97	98.1	98.2	98.5	92.4	93.6	93.9
0.99	98.1	98.2	98.4	92.5	93.6	94.1

Table 4: Analysis of PCA explained variance

\tau

. Image-level AUROC (I-AUROC) and PRO (%) are given for

k\!\in\!\{1,2,4\}

on both MVTec-AD and VisA datasets. Best results are in bold .

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Category	1-shot				2-shot				4-shot
Category	Img AUROC	Img AUPR	Pxl AUROC	PRO	Img AUROC	Img AUPR	Pxl AUROC	PRO	Img AUROC	Img AUPR	Pxl AUROC	PRO
Bottle	100.0 $\pm$ 0.0	100.0 $\pm$ 0.0	98.6 $\pm$ 0.0	96.5 $\pm$ 0.1	100.0 $\pm$ 0.0	100.0 $\pm$ 0.0	98.7 $\pm$ 0.0	96.7 $\pm$ 0.1	100.0 $\pm$ 0.0	100.0 $\pm$ 0.0	98.7 $\pm$ 0.0	96.7 $\pm$ 0.1
Cable	92.1 $\pm$ 1.4	95.5 $\pm$ 0.7	95.6 $\pm$ 0.2	90.3 $\pm$ 0.7	92.7 $\pm$ 1.2	96.2 $\pm$ 0.6	95.6 $\pm$ 0.2	90.6 $\pm$ 0.5	93.5 $\pm$ 0.7	96.7 $\pm$ 0.5	95.7 $\pm$ 0.2	90.9 $\pm$ 0.4
Capsule	88.1 $\pm$ 11.3	96.8 $\pm$ 3.9	98.0 $\pm$ 0.2	96.7 $\pm$ 0.5	92.7 $\pm$ 3.6	98.4 $\pm$ 0.9	98.2 $\pm$ 0.1	97.1 $\pm$ 0.2	93.4 $\pm$ 4.8	98.5 $\pm$ 1.2	98.3 $\pm$ 0.1	97.3 $\pm$ 0.2
Carpet	100.0 $\pm$ 0.0	100.0 $\pm$ 0.0	99.2 $\pm$ 0.0	98.2 $\pm$ 0.1	100.0 $\pm$ 0.0	100.0 $\pm$ 0.0	99.2 $\pm$ 0.0	98.2 $\pm$ 0.0	100.0 $\pm$ 0.0	100.0 $\pm$ 0.0	99.3 $\pm$ 0.0	98.3 $\pm$ 0.0
Grid	100.0 $\pm$ 0.0	100.0 $\pm$ 0.0	99.5 $\pm$ 0.0	98.1 $\pm$ 0.1	100.0 $\pm$ 0.0	100.0 $\pm$ 0.0	99.5 $\pm$ 0.0	98.2 $\pm$ 0.1	100.0 $\pm$ 0.0	100.0 $\pm$ 0.0	99.5 $\pm$ 0.0	98.3 $\pm$ 0.1

Appendix

Table 5: Detailed few-shot anomaly segmentation results of SubspaceAD on the MVTec AD dataset. We report mean Image AUROC (%), Image AUPR (%), Pixel AUROC (%), and Pixel PRO (%) results.

Category	1-shot				2-shot				4-shot
Category	Img AUROC	Img AUPR	Pxl AUROC	PRO	Img AUROC	Img AUPR	Pxl AUROC	PRO	Img AUROC	Img AUPR	Pxl AUROC	PRO
Candle	94.4 $\pm$ 0.8	94.2 $\pm$ 0.7	99.4 $\pm$ 0.0	97.9 $\pm$ 0.2	94.2 $\pm$ 0.5	93.8 $\pm$ 0.7	99.4 $\pm$ 0.0	98.1 $\pm$ 0.1	93.9 $\pm$ 0.9	93.4 $\pm$ 1.1	99.4 $\pm$ 0.0	98.1 $\pm$ 0.1
Capsules	96.5 $\pm$ 0.7	97.9 $\pm$ 0.4	98.5 $\pm$ 0.2	97.0 $\pm$ 0.3	96.8 $\pm$ 0.3	98.1 $\pm$ 0.2	98.6 $\pm$ 0.1	97.2 $\pm$ 0.3	97.0 $\pm$ 0.9	98.2 $\pm$ 0.6	98.6 $\pm$ 0.1	97.3 $\pm$ 0.3
Cashew	97.7 $\pm$ 0.7	98.9 $\pm$ 0.3	99.0 $\pm$ 0.0	98.5 $\pm$ 0.2	97.9 $\pm$ 1.3	99.0 $\pm$ 0.6	99.2 $\pm$ 0.1	98.6 $\pm$ 0.1	98.7 $\pm$ 0.8	99.4 $\pm$ 0.4	99.3 $\pm$ 0.1	98.7 $\pm$ 0.1
Chewing Gum	99.2 $\pm$ 0.1	99.6 $\pm$ 0.1	99.6 $\pm$ 0.0	96.0 $\pm$ 0.2	99.1 $\pm$ 0.1	99.6 $\pm$ 0.0	99.6 $\pm$ 0.0	96.3 $\pm$ 0.1	99.1 $\pm$ 0.0	99.5 $\pm$ 0.0	99.6 $\pm$ 0.0	96.2 $\pm$ 0.1
Fryum	97.0 $\pm$ 0.6	98.7 $\pm$ 0.3	96.5 $\pm$ 0.3	95.4 $\pm$ 0.5	97.9 $\pm$ 0.1	99.1 $\pm$ 0.1	96.8 $\pm$ 0.1	95.5 $\pm$ 0.3	98.1 $\pm$ 0.3	99.2 $\pm$ 0.1	97.0 $\pm$ 0.1	95.7 $\pm$ 0.3

Appendix

Table 6: Detailed few-shot anomaly segmentation results of SubspaceAD on the VisA dataset. We report mean Image AUROC (%), Image AUPR (%), Pixel AUROC (%), and Pixel PRO (%) results.

		MVTec-AD				VisA
		Image-level		Pixel-level		Image-level		Pixel-level
Setup	Method	AUROC	AUPR	AUROC	PRO	AUROC	AUPR	AUROC	PRO
1-shot	SubspaceAD 448 (DINOv3)	96.4	98.0	97.3	94.0	91.8	91.9	98.0	93.8
1-shot	SubspaceAD 672 (DINOv3)	96.2	97.9	97.3	94.2	92.8	92.9	98.0	94.7
2-shot	SubspaceAD 448 (DINOv3)	97.1	98.3	97.6	94.5	92.9	92.7	98.2	94.3
2-shot	SubspaceAD 672 (DINOv3)	96.8	98.2	97.5	94.7	93.7	93.6	98.1	95.1

Appendix

Table 7: Few-shot anomaly detection and localization results using the DINOv3-7B backbone. This backbone is consistently outperformed by DINOv2-G (see Table 1 in the main manuscript).

		MVTec-AD				VisA
		Image-level		Pixel-level		Image-level		Pixel-level
Setup	Method	AUROC	AUPR	AUROC	PRO	AUROC	AUPR	AUROC	PRO
Full-shot	AnomalyDINO	99.5	99.8	98.2	95.0	97.6	98.0	98.8	96.1
Full-shot	SubspaceAD (Ours)	99.2	99.6	98.2	95.6	98.2	98.3	99.1	96.9

Appendix

Table 8: Comparison of anomaly detection and localization performance on MVTec-AD and VisA in the full-shot setting. Best results are in bold , and second-best are underlined .

Method	Setting ( $K$ )	Scoring Time (ms / img)
AnomalyDINO	1-shot	$80.1$
AnomalyDINO	2-shot	$80.3$
AnomalyDINO	4-shot	$80.5$
SubspaceAD (Ours)	1–4 shot	74.1

Appendix

Table 9: Hardware-normalized scoring head latency (H100). We measure the time to process features after extraction. SubspaceAD latency is invariant to

K

, whereas memory-bank retrieval scales with the number of shots.

Method	Backbone	Params (M)	GPU	Time (ms / image)
WinCLIP [ 18 ]	CLIP ViT-B/16+	$150.0$	NVIDIA T4	$389$
AnomalyDINO [ 11 ]	DINOv2 ViT-S	$21.7$	NVIDIA A40	$60$
AnomalyDINO [ 11 ]	DINOv2 ViT-B	$85.8$	NVIDIA A40	$84$
AnomalyDINO [ 11 ]	DINOv2 ViT-L	$303.3$	NVIDIA A40	$141$
SubspaceAD (Ours) – DINOv2 Backbones (NVIDIA H100)
SubspaceAD (Ours)	DINOv2 ViT-S	$21.7$	NVIDIA H100	$36$

Appendix

Table 10: Inference time comparison with AnomalyDINO on MVTec-AD (1-shot, 448px ). Note that AnomalyDINO results are from the authors’ paper [ 11 ] , measured on an NVIDIA A40, while our measurements are obtained on an NVIDIA H100.

Figure 6: Additional qualitative results on VisA. Examples from six categories ( PCB1–4 , Chewing Gum , Fryum ). Rows show the input image, ground-truth mask, and our prediction. SubspaceAD accurately localizes both subtle texture anomalies and fine structural defects across diverse VisA domains.

Figure 7: Additional qualitative results on MVTec-AD. Examples from six categories ( Bottle , Cable , Metal Nut , Pill , Zipper , Grid ). Rows show the input image, ground-truth mask, and our prediction. SubspaceAD effectively localizes structural, positional, and surface anomalies across varied MVTec categories.

Figure 8: Qualitative failure cases. Each row shows the image, ground-truth mask, and our anomaly map. Examples include missed structural defects (e.g., missing transistor component), incorrect detection of cable swaps, and false positives from intra-class variability or background clutter.

Shots	Img AUROC	Img AUPR	Pxl AUROC	Pxl PRO
1-shot	89.1 $\pm$ 7.5	84.7 $\pm$ 5.7	82.3 $\pm$ 3.1	62.6 $\pm$ 2.0
2-shot	92.6 $\pm$ 3.6	88.5 $\pm$ 4.7	83.6 $\pm$ 0.9	63.7 $\pm$ 0.9
4-shot	93.4 $\pm$ 4.0	90.3 $\pm$ 4.3	84.0 $\pm$ 1.9	64.4 $\pm$ 1.3

Appendix

Table 11: Few-shot anomaly segmentation results of SubspaceAD on the MVTec-AD Transistor category using a rotation-agnostic protocol. We report mean Image AUROC (%), Image AUPR (%), Pixel AUROC (%), and Pixel PRO (%)

\pm

standard deviation across 5 seeds.

Explore similar work

DuoAD: Leveraging [CLS] Dual Characteristics for Training-Free Few-Shot Anomaly Detection
Jul 27, 2026Jyun-Ze Tang, Po-Han Huang, Ming-Ching Chang +2Few-Shot Anomaly Detection Anomaly Localization
Inventec Corporation, Taipei, Taiwan · University at Albany, State University of New York, NY, USA
Hypergraph-Enhanced Training-Free and Language-Free Few-Shot Anomaly Detection
May 11, 2026Guohuan Xie, Xin He, Dingying Fan +2Few-Shot Anomaly Detection Medical Image Anomaly Detection
Nankai University · Tianjin University of Technology · Tsinghua University
VisualAD: Language-Free Zero-Shot Anomaly Detection via Vision Transformer
Mar 9, 2026Yanning Hou, Peiyuan Li, Zirui Liu +4Zero-Shot Anomaly Detection Vision Foundation Model Adaptation
College of Intelligence Science and Technology, National University of Defense Technology, Changsha, China · School of Artificial Intelligence, Anhui University, Hefei, China · State Key Laboratory of Opto-Electronic Information Acquisition and Protection Technology, Anhui University, Hefei, China

Jul 27, 2026J/K move · Enter open · S save

DuoAD: Leveraging [CLS] Dual Characteristics for Training-Free Few-Shot Anomaly Detection

Jyun-Ze Tang, Po-Han Huang, Ming-Ching Chang, Chih-Fan Hsu+1

Inventec Corporation, Taipei, Taiwan · University at Albany, State University of New York, NY, USA

Few-Shot Anomaly Detection Anomaly Localization Anomaly Detection

Open paper AskPDF arXiv

Vision foundation models have enabled strong training-free anomaly detection (AD). However, most existing approaches rely primarily on independent local patch features, leaving the global contextual information encoded by Vision Transformers (ViTs) underexploited. In this work, we identify the dual characteristics of the ViT [CLS] token: its embedding provides anomaly-invariant global semantic representation, while its attention maps implicitly highlight spatially abnormal regions. Building on this observation, we propose a fully automated AD framework leveraging global context to remove manual tunings. Our framework introduces (1) an automatic augmentation selection strategy driven by [CLS]-level semantic consistency, and (2) an attention-guided feature reweighting mechanism that dynamically adjusts patch contributions according to [CLS] attention saliency. By integrating these components over multi-level features, our method achieves stable anomaly scoring and precise localization without training or parameter tuning. Under the one-shot setting, it achieves Image-AUC scores of 97.7%, 93.2%, and 84.5% on MVTec-AD, VisA, and Real-IAD. Using a single fixed configuration across categories, backbones, and datasets, the method establishes a new state-of-the-art for plug-and-play, training-free anomaly detection while maintaining strong robustness and practical scalability.

Abstract

Explore similar work

DuoAD: Leveraging [CLS] Dual Characteristics for Training-Free Few-Shot Anomaly Detection

Hypergraph-Enhanced Training-Free and Language-Free Few-Shot Anomaly Detection

VisualAD: Language-Free Zero-Shot Anomaly Detection via Vision Transformer