SPARC: Subspace Position-Aware Robust Few-Shot Calibration for Distribution-Shifted Industrial Anomaly Detection

Seokhee Han; Seungjun Chu; Mateusz Nowak; Peter Chin

cs.CVAug 19, 2026

SPARC: Subspace Position-Aware Robust Few-Shot Calibration for Distribution-Shifted Industrial Anomaly Detection

Authors: Seokhee Han, Seungjun Chu, Mateusz Nowak, Peter Chin

Organizations: Korea University

Abstract

Vision-based industrial anomaly detectors are calibrated on one distribution but may be deployed on another that differs in illumination, fixture placement, or sensor characteristics, sharply degrading an otherwise accurate detector. Adapting to the incoming lot is a natural response, but labeled anomalies are scarce. We therefore consider calibration using only a handful of verified-normal images available before scoring the rest of the lot. Existing fixes require backpropagation, detector-specific tuning, or choices about feature directions that few calibration samples cannot justify. We present SPARC, a few-shot calibration method that intercepts patch features between encoder and detector and removes a closed-form, spatially indexed estimate of deployment-time nuisance through per-cell subspace projection. It needs only $k \le 8$ verified-normal images and uses the algebraic saturation rank $r{=}k{-}1$ on the encoder's native patch grid. The correction requires no gradient or weight updates and works with memory-bank, density, prototype, and mutual detectors. On the shift-prone benchmarks, SPARC improves pooled Image AUROC and AU-PRO $_{0.3}$ for all seven detectors whose image scores depend on corrected patch features by $+13.8$ and $+3.5$ percentage points (pp), respectively; on benchmarks without engineered shift, the changes are small and mixed. Controls that give competing corrections the same calibration images attribute these gains to the per-cell subspace structure rather than the images alone. Further ablations support the saturation-rank choice and characterize sensitivity to backbone and calibration conditions.

Figures & tables

Figure 1: Overview of SPARC. (a) The pipeline consists of a frozen encoder, SPARC, and a frozen off-the-shelf detector. (b) At each cell, SPARC projects out an at-most- $(k{-}1)$ -dimensional disagreement subspace estimated from $k{\leq}8$ verified normals.

		MVTec AD 2			AeBAD-S			VisA			RAD			Pooled $\Delta$
Backbone	Method	base	+S	$\Delta$	base	+S	$\Delta$	base	+S	$\Delta$	base	+S	$\Delta$	shift prone	shift free
Image AUROC
WRN-50	PatchCore ( Roth et al., 2022 )	68.7	83.6	+14.8	64.1	66.3	+2.2	94.3	93.9	$-$ 0.3	92.1	96.2	+4.0	+10.6	+0.8
WRN-50	PaDiM ( Defard et al., 2021 )	59.9	81.6	+21.7	61.6	62.5	+1.0	89.3	89.6	+0.3	97.5	97.8	+0.3	+14.8	+0.3
WRN-50	SPADE ( Cohen and Hoshen, 2020 )	53.8	71.0	+17.2	38.3	43.6	+5.2	83.7	84.2	+0.5	28.9	34.7	+5.8	+13.2	+1.8
DINOv2-S/14	AnomalyDINO ( Damm et al., 2025 )	71.7	82.0	+10.3	71.7	73.5	+1.8	95.3	93.7	$-$ 1.6	97.2	97.0	$-$ 0.2	+7.5	$-$ 1.3

Table 1: Experimental results. Per benchmark, we report the uncorrected ( base ) and SPARC-corrected ( +S ) Image AUROC and AU-PRO 0.3 scores at

k{=}8

. All

\Delta

values are paired SPARC-minus-base differences computed before rounding and reported in percentage points (pp). The pooled columns aggregate category-level

\Delta

values across the two shift-prone benchmarks (MVTec AD 2 and AeBAD-S) and the two shift-free benchmarks (VisA and RAD), so each category receives equal weight. Bold marks the larger of base and +S within each dataset cell. AnomalyCLIP V and AnomalyCLIP M score images from the CLS token, so the per-cell correction leaves their Image AUROC unchanged. It still applies to their patch features and therefore affects AU-PRO.

Method	base	+S	$\Delta$
PatchCore	19.0	20.4	+1.3
PaDiM	11.3	17.1	+5.8
SPADE	18.4	19.5	+1.1
AnomalyDINO	37.0	38.8	+1.8
SubspaceAD	35.4	45.6	+10.2
WinCLIP	17.1	17.7	+0.7

Table 2: MVTec AD 2 localization results at low false-positive rates. AU-PRO 0.05 at

k{=}8

\Delta

is the paired SPARC-minus-base difference computed before rounding and reported in percentage points (pp).

Figure 2: Qualitative results with PatchCore. Base and SPARC share a common color scale within each column. Results use deployment-lot calibration with $k{=}8$ and seed $=0$ . Score maps are restored to native image coordinates. From left to right, the displayed categories are vial, rice, and can from MVTec AD 2, followed by candle, PCB4, and chewing gum from VisA.

Variant	$\Delta$ Image AUROC	$\Delta$ AU-PRO 0.3	MVTec AD 2 $\Delta$ AU-PRO 0.05
Shift-prone benchmarks (MVTec AD 2 and AeBAD-S)
CORAL few-shot	+1.8 $\pm$ 0.4	$-$ 2.8 $\pm$ 0.1	$-$ 2.5 $\pm$ 0.1
Calibration-only BN	$-$ 0.5 $\pm$ 0.4	+2.1 $\pm$ 0.2	+4.4 $\pm$ 0.1
Source-mixed BN	+1.5 $\pm$ 0.2	+4.0 $\pm$ 0.1	+6.1 $\pm$ 0.2
CORAL full-normal †	+3.0 $\pm$ 0.4	$-$ 2.7 $\pm$ 0.0	$-$ 2.6 $\pm$ 0.0
BN full-normal †	+0.4 $\pm$ 0.1	+2.3 $\pm$ 0.0	+4.7 $\pm$ 0.0

Table 3: Feature-level adaptation controls. All entries are paired changes in percentage points (pp). Values are macro-averaged over categories and methods within each seed and reported as mean

\pm

sample SD across five seeds. † Full-normal diagnostics use all ground-truth normal deployment images, including normals in the evaluation subset, and are nondeployable transductive controls.

Figure 3: Projection rank ablation. Panels (a) and (b) show

\Delta

Image AUROC on shift-prone MVTec AD 2 and shift-free VisA. Panel (c) shows

\Delta

AU-PRO 0.05 on MVTec AD 2, and panel (d) shows

\Delta

AU-PRO 0.3 on VisA. AnomalyCLIP appears only in the AU-PRO panels.

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Node	GPU configuration	VRAM	CPU configuration	Cores/Threads	System Memory
A	$2\times$ NVIDIA L40	48 GB	$2\times$ Intel Xeon Silver 4410Y	24/48	0.5 TB
B	$8\times$ NVIDIA RTX 6000 Ada Generation	48 GB	$2\times$ AMD EPYC 9124	32/64	1.5 TB
C	$6\times$ NVIDIA RTX 6000 Ada Generation	48 GB	$2\times$ AMD EPYC 9124	32/64	1.5 TB

Appendix

Table 4: Hardware specifications. System memory/cores show host RAM/CPUs; VRAM is per device.

Library or framework	Version
Python	3.12.3
PyTorch	2.7.0+cu128
Torchvision	0.22.0+cu128
xFormers	0.0.30
CUDA runtime/toolkit	12.8/12.8.93
cuDNN	9.7.1

Appendix

Table 5: Software environment. Versions were fixed across all experimental nodes.

	MVTec AD 2			AeBAD-S		VisA		RAD
Method	Image AUROC	AU-PRO 0.05	AU-PRO 0.3	Image AUROC	AU-PRO 0.3	Image AUROC	AU-PRO 0.3	Image AUROC	AU-PRO 0.3
FastRef	83.7	22.7	46.4	68.0	82.4	83.7	76.4	98.1	90.3
FastRecon	75.5	17.4	41.6	57.1	78.8	72.0	63.2	98.2	86.7
SubspaceAD + SPARC	85.8	45.6	69.3	88.1	92.7	98.2	95.9	96.4	94.6
AnomalyDINO + SPARC	82.0	38.8	63.7	73.5	86.7	93.7	92.0	97.0	90.5
PaDiM + SPARC	81.6	17.1	47.5	62.5	82.6	89.6	82.9	97.8	76.3

Appendix

Table 6: Query-conditioned detector comparison. At

k{=}8

, the table reports per-dataset Image AUROC and localization results; +SPARC denotes our correction layer.

		MVTec AD 2			AeBAD-S			VisA			RAD
Method	Budget	base	+C	$\Delta$	base	+C	$\Delta$	base	+C	$\Delta$	base	+C	$\Delta$
Image AUROC
SubspaceAD	CORAL-fs	70.9	71.0	+0.2	73.6	77.2	+3.6	98.1	97.6	-0.5	96.3	95.6	-0.7
SubspaceAD	CORAL full-normal	70.9	71.1	+0.2	73.6	77.3	+3.7	98.1	97.6	-0.5	96.3	95.6	-0.7
AnomalyDINO	CORAL-fs	71.7	72.5	+0.8	71.7	73.4	+1.7	95.3	95.3	+0.0	97.2	97.4	+0.2
AnomalyDINO	CORAL full-normal	71.7	72.7	+1.0	71.7	73.5	+1.9	95.3	95.4	+0.1	97.2	97.4	+0.2

Appendix

Table 7: CORAL results by detector and dataset. Matched-budget and full-normal diagnostic conditions use the same paired evaluation protocol.

Method	base	$\Delta C_{fs,k=2}$	$\Delta C_{fs,k=4}$	$\Delta C_{fs,k=8}$	$\Delta C_{full}$
Image AUROC
SubspaceAD	84.7	+0.4	+0.5	+0.7	+0.7
AnomalyDINO	84.0	+0.5	+0.6	+0.7	+0.8
PaDiM	77.1	-2.4	-2.2	-2.3	-2.3
PatchCore	79.8	-1.0	-0.7	-0.4	-0.4
WinCLIP	73.7	+0.6	+0.8	+0.7	+0.8

Appendix

Table 8: CORAL calibration budget. Few-shot estimates use

k\in\{2,4,8\}

. The full-normal diagnostic uses all ground-truth normal deployment images, including normals in the evaluation subset.

Variant	PatchCore	PaDiM	SPADE	Anomaly DINO	Subspace AD	MuSc	Anomaly CLIP V	Anomaly CLIP M	Pooled	Pooled $\Delta$
Image AUROC
base	82.4	78.1	60.8	85.5	86.5	66.6	72.8	76.7	76.2	0.0
mean-only	86.4	78.1	60.9	64.1	92.3	25.3	72.8	76.7	69.6	$-$ 6.6
SPARC (global subspace)	84.0	78.9	63.8	84.5	88.3	66.5	72.8	76.7	76.9	+0.8
SPARC (per-cell, ours)	87.3	84.6	67.5	87.9	92.9	76.2	72.8	76.7	80.8	+4.6
AU-PRO 0.3

Appendix

Table 10: Correction-mode ablation. base : no correction; mean-only : per-cell centroid subtraction; SPARC (global subspace) : a single rank-

(k{-}1)

subspace pooled over all cells; SPARC (per-cell) : our method. Bold marks the best entry per column; Pooled

\Delta

is the change vs. base. Image AUROC and AU-PRO 0.3 pool all 28 dataset–category pairs; AU-PRO 0.05 pools the eight MVTec AD 2 categories only. All results use

k{=}8

MVTec AD 2

VisA

Method

r{=}1

r{=}2

r{=}3

r{=}4

r{=}5

r{=}6

r{=}7

r{=}1

r{=}2

r{=}3

r{=}4

r{=}5

r{=}6

r{=}7

Image AUROC

\Delta

SubspaceAD

+10.2

+13.1

+14.1

+14.5

+14.8

+14.9

+0.1

AnomalyDINO

+2.7

+5.8

+8.0

+9.3

+9.9

+10.2

+10.3

-0.2

-0.6

-1.0

-1.2

-1.4

-1.5

-1.6

PaDiM

+12.3

+18.9

+21.1

+21.4

+21.6

+21.7

+0.2

+0.3

PatchCore

+5.0

+10.1

+13.7

+14.8

+15.3

+15.5

+14.8

+0.3

+0.0

-0.2

-0.1

-0.3

Appendix

Table 11: Projection-rank sensitivity. Paired changes as the retained rank approaches

k{-}1

	MVTec AD 2			AeBAD-S		VisA		RAD
Variant	Image AUROC	AU-PRO 0.05	AU-PRO 0.3	Image AUROC	AU-PRO 0.3	Image AUROC	AU-PRO 0.3	Image AUROC	AU-PRO 0.3
base (no correction)	68.7	19.0	42.8	64.1	85.1	94.3	87.8	92.1	73.2
$+\,$ SPARC (train-side only)	60.2	16.9	39.9	63.8	85.2	89.0	84.6	93.2	73.8
$+\,$ SPARC (test-side only)	82.9	18.6	42.6	64.6	85.6	86.8	83.3	96.4	77.3
$+\,$ SPARC (symmetric, ours)	83.6	20.4	45.7	66.3	86.3	93.9	86.5	96.2	77.9

Appendix

Table 12: Symmetric (train

{+}

test) versus one-sided calibration across all four benchmarks (

k{=}8

). We compare applying SPARC to training features only, test features only, or both. Dataset values are category macro-averages over five seeds; bold denotes the best result in each column.

MVTec AD 2

AeBAD-S

VisA

RAD

Method

k{=}2

k{=}4

k{=}8

k{=}2

k{=}4

k{=}8

k{=}2

k{=}4

k{=}8

k{=}2

k{=}4

k{=}8

Image AUROC

\Delta

SubspaceAD

+4.9

+9.0

+14.9

+8.2

+11.7

+14.5

+0.0

+0.1

+0.2

+0.1

AnomalyDINO

+1.8

+4.5

+10.3

+0.5

+0.8

+1.8

-0.2

-0.7

-1.6

-0.1

-0.2

PaDiM

+5.9

+11.6

+21.7

+0.2

+0.4

+1.0

+0.0

+0.2

+0.3

+0.1

+0.2

+0.3

PatchCore

+3.5

+7.8

+14.8

+2.4

+2.2

-0.3

-0.4

-0.3

+1.2

+2.7

+4.0

Appendix

Table 13: Calibration-size sensitivity. Paired changes for

k\in\{2,4,8\}

Method	0/8	1/8	2/8	3/8	4/8
Image AUROC $\Delta$
SubspaceAD	+14.9	+11.5	+8.5	+3.5	+0.5
AnomalyDINO	+10.3	+7.3	+4.2	+0.1	-3.0
PaDiM	+21.7	+18.7	+13.7	+7.5	+3.4
PatchCore	+14.8	+11.7	+8.0	+2.6	-1.7
MuSc	+33.1	+27.9	+22.0	+13.3	+6.2

Appendix

Table 14: Calibration contamination. Paired changes on MVTec AD 2 as normal calibration images are replaced by same-category anomalies.

	MVTec AD 2			AeBAD-S			VisA			RAD
Backbone	base	+S	$\Delta$	base	+S	$\Delta$	base	+S	$\Delta$	base	+S	$\Delta$
Image AUROC
clip_b	64.5	77.7	+13.2	61.5	60.7	-0.8	90.2	84.1	-6.0	90.5	74.2	-16.3
convnext_b	61.8	78.3	+16.6	68.3	67.0	-1.4	83.6	83.5	-0.1	96.6	96.7	+0.1
dinov2_b	66.6	82.7	+16.2	61.5	67.1	+5.6	94.5	92.7	-1.8	95.6	96.1	+0.5
dinov3_b	65.2	83.0	+17.8	76.6	76.6	-0.0	95.4	94.1	-1.3	96.5	96.9	+0.3

Appendix

Table 15: Backbone sensitivity. PatchCore results with eight off-the-shelf feature encoders at

k{=}8

	Center crop (main)			Full frame (control)
Method	base	+S	$\Delta$	base	+S	$\Delta$
Image AUROC
AnomalyDINO	$95.3\mathbin{\pm}0.1$	$93.7\mathbin{\pm}0.2$	$-1.6\mathbin{\pm}0.1$	$95.4\mathbin{\pm}0.1$	$93.7\mathbin{\pm}0.2$	$-1.7\mathbin{\pm}0.1$
PaDiM	$89.3\mathbin{\pm}0.1$	$89.6\mathbin{\pm}0.2$	$+0.3\mathbin{\pm}0.1$	$86.3\mathbin{\pm}0.3$	$86.5\mathbin{\pm}0.3$	$+0.2\mathbin{\pm}0.0$
PatchCore	$94.3\mathbin{\pm}0.2$	$93.9\mathbin{\pm}0.2$	$-0.3\mathbin{\pm}0.3$	$93.7\mathbin{\pm}0.1$	$93.0\mathbin{\pm}0.2$	$-0.7\mathbin{\pm}0.3$
SPADE	$83.7\mathbin{\pm}0.2$	$84.2\mathbin{\pm}0.5$	$+0.5\mathbin{\pm}0.6$	$83.4\mathbin{\pm}0.3$	$84.1\mathbin{\pm}0.5$	$+0.7\mathbin{\pm}0.4$

Appendix

Table 16: VisA input-geometry sensitivity. Center crop (main) versus aspect-ratio-preserving full-frame resizing with padding (control) at

k{=}8

. Split, settings, calibration draws, and held-out images are fixed. Category macro-averages are mean

\pm

sample SD over five paired seeds.

\Delta

is SPARC minus base before rounding (pp).

		Image AUROC			AU-PRO 0.05			AU-PRO 0.3
Backbone	Method	base	+S	paired $\Delta$	base	+S	paired $\Delta$	base	+S	paired $\Delta$
WRN-50	PatchCore ( Roth et al., 2022 )	68.3 $\pm$ 2.2	68.9 $\pm$ 2.2	+0.6 $\pm$ 2.1	18.9 $\pm$ 0.1	20.6 $\pm$ 0.3	+1.7 $\pm$ 0.2	42.6 $\pm$ 0.3	43.0 $\pm$ 0.2	+0.4 $\pm$ 0.2
WRN-50	PaDiM ( Defard et al., 2021 )	60.0 $\pm$ 2.9	62.0 $\pm$ 2.7	+2.0 $\pm$ 0.6	11.3 $\pm$ 0.1	18.4 $\pm$ 0.1	+7.0 $\pm$ 0.1	40.9 $\pm$ 0.2	44.7 $\pm$ 0.3	+3.8 $\pm$ 0.2
WRN-50	SPADE ( Cohen and Hoshen, 2020 )	54.3 $\pm$ 1.0	57.3 $\pm$ 2.1	+3.1 $\pm$ 1.3	18.3 $\pm$ 0.0	18.2 $\pm$ 0.3	$-$ 0.1 $\pm$ 0.3	41.9 $\pm$ 0.3	42.7 $\pm$ 0.3	+0.8 $\pm$ 0.2
DINOv2-S/14	AnomalyDINO ( Damm et al., 2025 )	71.5 $\pm$ 0.4	72.2 $\pm$ 0.8	+0.7 $\pm$ 0.5	36.8 $\pm$ 0.0	38.0 $\pm$ 0.2	+1.2 $\pm$ 0.2	59.9 $\pm$ 0.1	61.7 $\pm$ 0.7	+1.9 $\pm$ 0.7
DINOv2-G	SubspaceAD ( Lendering et al., 2026 )	71.9 $\pm$ 0.9	77.3 $\pm$ 1.8	+5.4 $\pm$ 1.0	35.3 $\pm$ 0.1	44.4 $\pm$ 1.1	+9.0 $\pm$ 1.0	61.2 $\pm$ 0.1	67.9 $\pm$ 0.7	+6.7 $\pm$ 0.6

Appendix

Table 17: Cross-scene transfer on MVTec AD 2. Calibration-scene captures are excluded; values are category macro-averages reported as mean

\pm

sample SD over five paired seeds.

		base		mean-only		global subspace		per-cell SPARC (ours)		SPARC + mean
Detector	Metric	Score	$\Delta$	Score	$\Delta$	Score	$\Delta$	Score	$\Delta$	Score	$\Delta$
SubspaceAD	Image AUROC	84.7	+0.0	91.5	+6.8	87.2	+2.5	92.1	+7.4	92.2	+7.5
	AU-PRO 0.3	85.2	+0.0	87.6	+2.5	86.4	+1.2	88.1	+3.0	88.2	+3.0
	AU-PRO 0.05	35.4	+0.0	44.5	+9.1	39.0	+3.6	45.6	+10.2	45.7	+10.2
AnomalyDINO	Image AUROC	84.0	+0.0	64.7	-19.3	82.9	-1.1	86.5	+2.6	52.0	-31.9
	AU-PRO 0.3	81.4	+0.0	35.5	-45.9	76.9	-4.5	83.2	+1.8	14.9	-66.5

Appendix

Table 18: Mean-removal variants. Detector-level results for subspace and mean corrections at

k{=}8

Dataset	DC energy	Subspace energy	DC fraction
MVTec AD 2	43.77	154.12	0.221
AeBAD-S	76.84	232.90	0.248
VisA	16.67	117.31	0.124
RAD	65.69	184.73	0.262

Appendix

Table 19: Disagreement-energy decomposition. The DC fraction is the share of the combined DC and subspace energy attributable to the per-cell mean offset.

Method	Deploy $\Delta$	Train $\Delta$	Difference	Raw $p$	Holm $p$
Image AUROC
PatchCore	+10.6	-2.5	+13.1	$<$ .001 ∗∗∗	.003 ∗∗
PaDiM	+14.8	-0.0	+14.8	$<$ .001 ∗∗∗	.003 ∗∗
SPADE	+13.2	+1.2	+12.0	$<$ .001 ∗∗∗	.003 ∗∗
AnomalyDINO	+7.5	-0.3	+7.7	.001 ∗∗	.003 ∗∗
SubspaceAD	+14.8	-0.2	+15.0	$<$ .001 ∗∗∗	.003 ∗∗

Appendix

Table 20: Calibration-source control. One-sided Wilcoxon signed-rank tests compare deployment-lot and training-pool calibration under the same paired protocol. Holm correction is applied jointly across six detectors and two primary metrics.

	Anomaly–subspace alignment		Anomaly–normal contrast
Dataset	$\rho_{\mathrm{anom}}$	$\rho_{\mathrm{anom}}/(7/1536)$	base	SPARC	$\Delta$	Rel. $\Delta$
MVTec AD 2	0.119	26.2 $\times$	0.749	1.139	+0.390	+52.1%
AeBAD-S	0.064	14.1 $\times$	1.576	1.776	+0.200	+12.7%
VisA	0.195	42.9 $\times$	1.338	1.342	+0.004	+0.3%
RAD	0.072	15.8 $\times$	1.693	1.822	+0.130	+7.7%

Appendix

Table 21: Anomaly–subspace alignment and anomaly–normal contrast.

\Delta

is SPARC minus base. Ratios and changes are computed before rounding; displayed entries are rounded independently. VisA uses the center-crop geometry of the primary experiment; the full-frame control is reported in Table 16 . Feature setup, the

7/1536

reference, and the contrast definition are given in Section G.3 .

Method	$N$	Mean $\Delta$	Median $\Delta$	Raw $p$	Holm $p$
Image AUROC
PatchCore	12	+10.62	+6.86	.001 ∗∗	.006 ∗∗
PaDiM	12	+14.78	+13.22	$<$ .001 ∗∗∗	.003 ∗∗
SPADE	12	+13.18	+11.14	$<$ .001 ∗∗∗	.003 ∗∗
AnomalyDINO	12	+7.47	+6.87	$<$ .001 ∗∗∗	.003 ∗∗
SubspaceAD	12	+14.77	+13.36	$<$ .001 ∗∗∗	.003 ∗∗

Appendix

Table 22: Primary shift-prone significance tests. One-sided Wilcoxon signed-rank tests over seed-averaged category changes on MVTec AD 2 and AeBAD-S. Holm correction is applied jointly across the seven detectors and two primary metrics.

	MVTec AD 2		AeBAD-S		VisA		RAD
Method	Mean $\Delta$	Seed SD	Mean $\Delta$	Seed SD	Mean $\Delta$	Seed SD	Mean $\Delta$	Seed SD
Image AUROC
SubspaceAD	+14.9	0.8	+14.5	1.0	+0.1	0.1	+0.1	0.1
AnomalyDINO	+10.3	1.4	+1.8	0.4	-1.6	0.1	-0.2	0.1
PaDiM	+21.7	2.9	+1.0	0.5	+0.3	0.1	+0.3	0.1
PatchCore	+14.8	2.0	+2.2	1.3	-0.3	0.3	+4.0	0.4

Appendix

Table 23: Calibration-draw variability. Mean paired changes and sample standard deviations over five seeds.

Category	Image AUROC	AU-PRO 0.05	AU-PRO 0.3
MVTec AD 2
can	55.7/64.8	4.9/10.5	26.3/32.9
fabric	67.8/73.2	3.5/3.4	18.3/18.0
fruit_jelly	78.1/94.5	32.4/35.7	60.5/63.5
rice	75.4/81.8	17.9/22.0	40.3/43.0
sheet_metal	91.2/92.0	9.5/9.5	26.2/25.5

Appendix

Table 24: Per-category few-shot baselines. Entries are FastRecon/FastRef at

k{=}8

Category	Variant	PatchCore	PaDiM	SPADE	AnomalyDINO	SubspaceAD	WinCLIP	MuSc	AnomalyCLIP V	AnomalyCLIP M
can	base	49.7	49.7	50.0	52.5	50.0	24.2	53.0	57.2	55.2
can	$+\,$ SPARC	60.2	58.9	63.1	63.5	68.3	44.8	69.8	57.2	55.2
fabric	base	52.0	28.2	32.5	56.0	64.8	58.8	34.9	58.3	75.0
fabric	$+\,$ SPARC	73.2	65.1	63.9	70.5	77.4	86.4	71.4	58.3	75.0
fruit_jelly	base	89.2	79.2	70.8	86.5	83.4	88.1	38.5	66.1	67.3
fruit_jelly	$+\,$ SPARC	96.8	96.4	95.4	98.9	99.8	92.9	93.1	66.1	67.3

Appendix

Table 25: Per-category Image AUROC on MVTec AD 2. Results use

k{=}8

Category	Variant	PatchCore	PaDiM	SPADE	AnomalyDINO	SubspaceAD	WinCLIP	MuSc	AnomalyCLIP V	AnomalyCLIP M
bolt	base	79.1	92.5	13.0	95.1	95.6	75.2	75.8	76.1	87.5
bolt	$+\,$ SPARC	90.8	93.9	15.8	94.9	96.0	81.8	67.5	76.1	87.5
ribbon	base	96.6	99.2	77.0	98.2	96.6	85.1	81.0	98.7	92.7
ribbon	$+\,$ SPARC	97.4	99.1	84.8	97.9	96.7	86.1	78.7	98.7	92.7
sponge	base	95.5	99.2	7.2	97.2	96.0	68.1	55.4	43.8	36.6
sponge	$+\,$ SPARC	97.6	99.2	13.1	97.0	96.0	72.9	59.1	43.8	36.6

Appendix

Table 29: RAD per-category Image AUROC (

k{=}8

Figure 4: SPARC-corrected qualitative localization on MVTec AD 2. All anomaly maps use SPARC-corrected features; the nine detector columns are min-max normalized independently per image and detector and rendered with a shared turbo colormap. Quantitative results are reported in the main tables. For each category, the displayed test image is the one selected by lower median GT-mask area; filenames are sorted in ascending order on ties ( $k{=}8$ , seed=0).

Figure 5: SPARC-corrected qualitative localization on VisA. All anomaly maps use SPARC-corrected features; the nine detector columns are min-max normalized independently per image and detector and rendered with a shared turbo colormap. Quantitative results are reported in the main tables. For each category, the displayed test image is the one selected by lower median GT-mask area; filenames are sorted in ascending order on ties ( $k{=}8$ , seed=0). (page 1 of 2)

Figure 6: SPARC-corrected qualitative localization on VisA. (page 2 of 2)

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

Feb 26, 2026

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

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.

Camile Lendering, Erkut Akdag, Egor Bondarev

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

Few-Shot Anomaly Detection Anomaly Detection Industrial Anomaly Detection Reconstruction-Based Anomaly Detection

Ask PDF

Oct 4, 2026

Anchor and Adapt: Asymmetric Prompt Adaptation for Few-Shot Industrial Anomaly Detection

In few-shot industrial anomaly detection, the few normal target images provide no direct defect supervision, making anomaly prompts difficult to learn from these samples alone. Some vision-language methods therefore use manually specified descriptions to supply explicit anomaly semantics. However, constructing these descriptions requires product-specific effort, and their effectiveness depends on prompt selection. We propose Anchor and Adapt, a two-stage prompt learning framework that separates the acquisition of anomaly semantics from adaptation to target normal appearance. Stage I learns transferable normal and abnormal anchors from annotated auxiliary data. Stage II keeps these anchors fixed and adapts an additional normal branch using the few target normal samples. The inherited and adapted normal branches jointly characterize target normality, with text-anchor regularization encouraging consistency with the generic normal prior and separation from the abnormal anchors. This design retains learned anomaly knowledge while reducing dependence on category-specific anomaly templates, without requiring synthetic anomaly generation. Cross-dataset experiments between MVTec-AD and VisA under 1-, 2-, and 4-shot settings demonstrate competitive detection and localization performance. Controlled ablations assess the roles of transferred anchors, asymmetric adaptation, dual-normal representations, and anchor regularization.

Mengyang Zhao, Teng Fu, Haiyang Yu +3

Shanghai Key Lab of Intelligent Information Processing and College of Computer Science and Artificial Intelligence, Fudan University, Shanghai 200433, China

Prompt Learning for VLMs Few-Shot Anomaly Detection Industrial Anomaly Detection VLM Adaptation Anomaly Localization

Ask PDF

Aug 29, 2026

NFAD: Nuisance-Filtered Anomaly Detection Under Distribution Shift

Recent advances in anomaly detection (AD) for industrial inspection have pushed performance on standard benchmarks toward saturation. However, strong benchmark performance does not necessarily translate to real-world deployment, as these benchmarks are primarily collected under controlled acquisition conditions. Changes in illumination, background, viewpoint, and other environmental factors can shift normal samples away from the learned normal distribution and cause false anomaly responses. We address AD under such distribution shifts by explicitly modeling nuisance variation from changing imaging conditions in feature space. Without anomaly labels or target-domain data, our Nuisance-Filtered Anomaly Detection (NFAD) framework estimates a nuisance subspace from matched feature displacements induced by content-preserving perturbations and suppresses its contribution to anomaly residuals at inference. The same subspace supports two complementary branches: full projection for image-level detection and selective suppression for pixel-level localization, preserving evidence of localized defects. On AeBAD-S, a benchmark specifically designed for AD under acquisition shifts, NFAD achieves 91.0% image-level AUROC, establishing a new state of the art. Notably, this robustness does not come at the expense of conventional AD performance: NFAD remains competitive on standard benchmarks that do not explicitly evaluate distribution shift, including VisA, Real-IAD, and MVTec AD. These results show that explicitly suppressing such nuisance variation improves AD under distribution shift while preserving strong performance in standard settings.

Dat Cao, Son Nghiem, Phan Nguyen +2

Korea Advanced Institute of Science and Technology (KAIST), South Korea · Sony Computer Science Laboratories (Sony CSL), Japan · The University of Tokyo, Japan +1

Anomaly Detection Industrial Anomaly Detection Anomaly Localization Invariant Representation Learning

Ask PDF

Abstract

Industrial Anomaly Detection

Explore similar work

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

Anchor and Adapt: Asymmetric Prompt Adaptation for Few-Shot Industrial Anomaly Detection

NFAD: Nuisance-Filtered Anomaly Detection Under Distribution Shift