Flownomaly: Single-Step Conditional-Mean Recovery for Few-Shot Anomaly Detection

Bernardo Perrone Ribeiro; Erkut Akdag; Egor Bondarev

cs.CVOct 23, 2026

Flownomaly: Single-Step Conditional-Mean Recovery for Few-Shot Anomaly Detection

Authors: Bernardo Perrone Ribeiro, Erkut Akdag, Egor Bondarev

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

Abstract

Few-shot industrial anomaly detection requires modeling category-specific normality from only a handful of defect-free images. Common approaches tie inference to stored exemplars, impose rigid linear structure, or risk reproducing anomalous content. Flownomaly instead casts the problem as normal-counterpart recovery in frozen DINOv2 feature space. Given normal supports, a velocity field is learned through conditional flow matching over frozen patch tokens to transport Gaussian noise toward normal features. Along straight paths, a single recovery step using one evaluation of the optimal field yields the posterior mean of the clean normal token without trajectory integration. The operating time governs how strongly recovery modifies a query token. With no anomalous validation samples available for tuning, a normal-only Jacobian gain-to-noise criterion selects this time from the supports alone. Squared residuals between query tokens and their recovered counterparts yield anomaly signals for both dense localization and image-level detection. Across 1-, 2-, and 4-shot benchmarks on MVTec-AD and VisA, together with sample-shot evaluations on Real-IAD, Flownomaly achieves state-of-the-art performance against recent category-specific methods. In the 1-shot setting, image- and pixel-level AUROC reach 97.8% and 98.0% on MVTec-AD and 95.3% and 98.7% on VisA, with a 2.1-point image-level AUROC improvement over the strongest baseline on VisA. Code is available at https://github.com/b-rbmp/Flownomaly.

Figures & tables

Figure 1: Conceptual illustration of Flownomaly. A query patch token is (1) perturbed with Gaussian noise and (2) recovered using a single evaluation of the velocity field fitted to the normal supports. Tokens consistent with normal appearance undergo little displacement (left), whereas less compatible tokens tend to undergo larger changes toward the normal support features (right). The squared distance between a query token and its recovered counterpart defines the anomaly score.

Figure 2: Overview of Flownomaly. (a) Disjoint augmented views of the $k$ normal supports are encoded into standardized multi-depth DINOv2 tokens (Sec. 3.1 ): fitting views train the velocity field $v_{\theta}$ via conditional flow matching (Sec. 3.2 ), validation views select the checkpoint $v_{\widehat{\theta}}$ , and calibration views determine the operating time $t^{\star}$ via a normal-only criterion (Sec. 3.5 ). (b) At inference, query tokens perturbed with Gaussian noise at $t^{\star}$ are recovered in one batched evaluation of $v_{\widehat{\theta}}$ (Sec. 3.3 ). Squared residuals to recovered counterparts form the anomaly score map, which is smoothed and upsampled for dense localization (Sec. 3.4 ).

Setup	Method	MVTec-AD						VisA
		Image-level			Pixel-level			Image-level			Pixel-level
		AUROC	AUPR	$F_{1,\max}$	AUROC	AUPRO	$F_{1,\max}$	AUROC	AUPR	$F_{1,\max}$	AUROC	AUPRO	$F_{1,\max}$
1-shot	PatchCore [ 27 ]	83.4	92.2	90.5	92.0	79.7	50.4	79.9	82.8	81.7	95.4	80.5	38.0
	WinCLIP [ 16 ]	93.1	96.5	93.7	95.2	87.1	55.9	83.8	85.1	83.1	96.4	85.1	41.3
	PromptAD [ 20 ]	94.6	97.1	95.2	95.9	87.9	53.2	86.9	88.4	82.9	96.7	85.1	35.0
	AnomalyDINO [ 6 ]	96.6	98.2	95.8	96.8	92.7	60.2	87.4	89.0	84.3	97.8	92.5	45.1

Table 1: Few-shot anomaly detection performance on MVTec-AD and VisA for category-specific methods. Flownomaly is reported as mean

\pm

standard deviation over 3 seeds. All values are percentages. Bold and underlined indicate the best and second-best results, respectively.

Figure 3: Qualitative comparison on MVTec-AD and VisA under the 1-shot setting. Columns show the query image, ground-truth mask, and anomaly maps from the open-source category-specific baselines AnomalyDINO [ 6 ] and SubspaceAD [ 18 ] , followed by Flownomaly. Flownomaly delineates defects more sharply, with fewer spurious background activations and closer agreement with the ground-truth masks on both datasets. Additional qualitative results and representative failure modes are provided in Suppl. D and B.4 , respectively.

1-sample shot (5 views)
	Image-level			Pixel-level
Method	AUROC	AUPR	$F_{1,\max}$	AUROC	AUPRO	$F_{1,\max}$
PatchCore [ 27 ]	$79.5_{\scriptstyle\pm 0.1}$	$77.7_{\scriptstyle\pm 0.1}$	$71.5_{\scriptstyle\pm 0.0}$	$96.9_{\scriptstyle\pm 0.0}$	$84.3_{\scriptstyle\pm 0.1}$	$30.7_{\scriptstyle\pm 0.2}$
PromptAD [ 20 ]	$79.4_{\scriptstyle\pm 0.2}$	$77.2_{\scriptstyle\pm 0.4}$	$71.4_{\scriptstyle\pm 0.1}$	$95.9_{\scriptstyle\pm 0.1}$	$81.7_{\scriptstyle\pm 0.1}$	$30.8_{\scriptstyle\pm 0.2}$
AnomalyDINO [ 6 ]	$82.5_{\scriptstyle\pm 0.1}$	$78.3_{\scriptstyle\pm 0.4}$	$73.9_{\scriptstyle\pm 0.1}$	$97.6_{\scriptstyle\pm 0.0}$	$91.7_{\scriptstyle\pm 0.2}$	$37.0_{\scriptstyle\pm 0.5}$
SubspaceAD [ 18 ]	$85.2_{\scriptstyle\pm 0.3}$	$81.6_{\scriptstyle\pm 0.4}$	$76.1_{\scriptstyle\pm 0.0}$	$98.8_{\scriptstyle\pm 0.0}$	$\bm{97.1}_{\scriptstyle\pm 0.1}$	$42.7_{\scriptstyle\pm 0.2}$

Table 2: Sample-shot anomaly detection and localization on Real-IAD. Results are percentages reported as mean

\pm

standard deviation over 3 seeds. Bold and underlined indicate the best and second-best results, respectively.

	MVTec-AD		VisA		Latency (ms/image)
NFE	I-AUROC	P-AUROC	I-AUROC	P-AUROC	Velocity	Total model
1	$97.8_{\scriptstyle\pm 0.9}$	$98.0_{\scriptstyle\pm 0.1}$	$95.3_{\scriptstyle\pm 0.7}$	$98.7_{\scriptstyle\pm 0.0}$	2.2	60.9
2	$97.8_{\scriptstyle\pm 1.0}$	$98.1_{\scriptstyle\pm 0.1}$	$95.2_{\scriptstyle\pm 0.7}$	$98.7_{\scriptstyle\pm 0.0}$	3.7	62.4
5	$97.8_{\scriptstyle\pm 1.0}$	$98.1_{\scriptstyle\pm 0.1}$	$95.1_{\scriptstyle\pm 0.8}$	$98.7_{\scriptstyle\pm 0.0}$	8.0	66.7
10	$97.8_{\scriptstyle\pm 1.0}$	$98.1_{\scriptstyle\pm 0.1}$	$95.1_{\scriptstyle\pm 0.8}$	$98.7_{\scriptstyle\pm 0.0}$	15.1	73.8
20	$97.8_{\scriptstyle\pm 1.0}$	$98.1_{\scriptstyle\pm 0.1}$	$95.1_{\scriptstyle\pm 0.8}$	$98.7_{\scriptstyle\pm 0.0}$	29.2	87.9

Table 3: Effect of the number of velocity-field evaluations (NFE) in the 1-shot setting.

Figure 4: Operating-time robustness on MVTec-AD and VisA in the 1-shot setting. Curves show mean category-wise image-level regret; lower is better. The dashed lines mark the median category-wise JGN selection at

t=0.65

	MVTec-AD		VisA
Scoring rule	I-AUROC	P-AUROC	I-AUROC	P-AUROC
Recovery residual	$\bm{97.8}_{\scriptstyle\pm 0.9}$	$\bm{98.0}_{\scriptstyle\pm 0.1}$	$\bm{95.3}_{\scriptstyle\pm 0.7}$	$\bm{98.7}_{\scriptstyle\pm 0.0}$
Velocity norm	$96.9_{\scriptstyle\pm 0.9}$	$97.3_{\scriptstyle\pm 0.1}$	$94.3_{\scriptstyle\pm 0.2}$	$98.4_{\scriptstyle\pm 0.0}$
Cosine residual	$96.7_{\scriptstyle\pm 0.6}$	$97.9_{\scriptstyle\pm 0.1}$	$90.2_{\scriptstyle\pm 0.4}$	$97.8_{\scriptstyle\pm 0.2}$

Table 4: Effect of the scoring rule (1-shot setting), with all other factors fixed. Best results are in bold .

		MVTec-AD		VisA
Backbone	Latency (ms)	I-AUROC	P-AUROC	I-AUROC	P-AUROC
DINOv2-S	18.1	$94.7_{\scriptstyle\pm 1.3}$	$97.7_{\scriptstyle\pm 0.1}$	$91.8_{\scriptstyle\pm 1.0}$	$98.0_{\scriptstyle\pm 0.1}$
DINOv2-B	25.5	$96.4_{\scriptstyle\pm 1.0}$	$97.8_{\scriptstyle\pm 0.1}$	$93.4_{\scriptstyle\pm 0.8}$	$98.3_{\scriptstyle\pm 0.0}$
DINOv2-L	51.9	$97.3_{\scriptstyle\pm 0.9}$	$97.9_{\scriptstyle\pm 0.1}$	$94.3_{\scriptstyle\pm 0.8}$	$98.5_{\scriptstyle\pm 0.0}$
DINOv2-g	60.9	$\bm{97.8}_{\scriptstyle\pm 0.9}$	$\bm{98.0}_{\scriptstyle\pm 0.1}$	$\bm{95.3}_{\scriptstyle\pm 0.7}$	$\bm{98.7}_{\scriptstyle\pm 0.0}$

Table 5: Backbone-scale ablation in the 1-shot setting. Best results are in bold .

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1 : Recovery relative to the support-derived normal token set. Left: recovery residual versus pre-recovery support distance, with both quantities normalized by their category-wise normal-token medians and aggregated equally across categories. Middle:

d_{\mathrm{after}}/d_{\mathrm{before}}

, where values below one indicate contraction toward the normal support features. Right: cosine alignment between the recovery update and the direction toward the nearest support token, where positive values indicate movement toward the normal support features. Bars and error bars report the median and interquartile range across categories. Ground-truth masks serve only for post-hoc stratification.

		Regret (pp)
Dataset	$t^{\star}$ range	Mean	Median	Maximum
MVTec-AD	$0.60$ – $0.75$	0.15	0.06	0.82
VisA	$0.65$ – $0.70$	0.54	0.14	4.82

Appendix

Table S1 : Retrospective image-AUROC regret of the normal-only JGN selection, measured in percentage points relative to the category-wise test-label oracle. Regret statistics are computed over categories for seed 42; the

t^{\star}

column reports the range of per-category selections across all three seeds.

MVTec-AD			VisA
Category	$t^{\star}$	Regret (pp)	Category	$t^{\star}$	Regret (pp)
bottle	0.75	0.08	candle	0.70	0.09
cable	0.65	0.82	capsules	0.65	0.36
capsule	0.65	0.08	cashew	0.65	0.02
carpet	0.70	0.00	chewinggum	0.65	0.04
grid	0.75	0.00	fryum	0.65	0.23

Appendix

Table S2 : Category-level operating-time diagnostics for seed 42. Regret is the image-AUROC difference, in percentage points, from the retrospective test-label oracle.

Figure S2 : Distribution of category-specific operating times selected by JGN, pooled across three seeds. Each observation represents one category–seed pair.

	MVTec-AD		VisA
Configuration	I-AUROC	P-AUROC	I-AUROC	P-AUROC	Median $t^{*}$ (MVTec-AD/VisA)
Default: 40–80%, standardized	$97.8_{\scriptstyle\pm 0.9}$	$98.0_{\scriptstyle\pm 0.1}$	$95.3_{\scriptstyle\pm 0.7}$	$98.7_{\scriptstyle\pm 0.0}$	0.65/0.65
Single block at 80% depth	$97.6_{\scriptstyle\pm 1.2}$	$98.0_{\scriptstyle\pm 0.0}$	$92.0_{\scriptstyle\pm 0.7}$	$98.9_{\scriptstyle\pm 0.0}$	0.70/0.70
Shallow blocks at 20–40% depth	$85.7_{\scriptstyle\pm 0.9}$	$95.9_{\scriptstyle\pm 0.1}$	$86.0_{\scriptstyle\pm 1.7}$	$93.5_{\scriptstyle\pm 0.3}$	0.60/0.60
Without layer standardization	$97.7_{\scriptstyle\pm 1.3}$	$98.2_{\scriptstyle\pm 0.1}$	$94.0_{\scriptstyle\pm 0.9}$	$98.8_{\scriptstyle\pm 0.0}$	0.85/0.85
Without rotation-derived views †	$93.8_{\scriptstyle\pm 1.4}$	$97.4_{\scriptstyle\pm 0.1}$	$93.3_{\scriptstyle\pm 0.5}$	$98.5_{\scriptstyle\pm 0.0}$	0.60/0.60

Appendix

Table S3 : Feature and support-view ablations in the 1-shot setting. AUROC results are mean

\pm

standard deviation over three seeds. The final column reports the median selected

t^{\star}

over categories for MVTec-AD/VisA under seed 42. † Disabling rotations also removes diversity from the support-derived validation and calibration pools.

Figure S3 : Failure modes where the predicted anomaly map remains essentially empty. Columns show the query image, ground-truth mask, and Flownomaly anomaly map. Top: Logical anomaly with swapped wires (MVTec-AD cable ). Bottom: Missing component revealing previously seen background (VisA pcb3 ).

		MVTec-AD		VisA
Backbone	Res.	I-AUROC	P-AUROC	I-AUROC	P-AUROC
DINOv2-g	$672^{2}$	$97.8_{\scriptstyle\pm 0.9}$	$98.0_{\scriptstyle\pm 0.1}$	$95.3_{\scriptstyle\pm 0.7}$	$98.7_{\scriptstyle\pm 0.0}$
DINOv3-7B	$672^{2}$	$96.1_{\scriptstyle\pm 1.0}$	$98.0_{\scriptstyle\pm 0.1}$	$94.8_{\scriptstyle\pm 0.1}$	$98.5_{\scriptstyle\pm 0.0}$
DINOv3-7B	$768^{2}$	$96.0_{\scriptstyle\pm 1.6}$	$97.9_{\scriptstyle\pm 0.1}$	$94.9_{\scriptstyle\pm 0.2}$	$98.5_{\scriptstyle\pm 0.0}$

Appendix

Table S4 : DINOv3 vs. DINOv2 backbone ablation in the 1-shot setting across MVTec-AD and VisA. DINOv3 (

16\times 16

patch size) is evaluated at

672\times 672

(

42\times 42

token grid) and at

768\times 768

to match DINOv2’s

48\times 48

token grid.

	Image-level			Pixel-level
Method	AUROC	AUPR	$F_{1,\max}$	AUROC	AUPRO	$F_{1,\max}$
PatchCore [ 27 ]	99.2	99.8	98.8	98.1	92.6	59.3
AnomalyDINO [ 6 ]	99.5	99.8	99.0	98.2	95.0	64.3
SubspaceAD [ 18 ]	99.2	99.6	98.3	98.2	95.6	64.7
ANoCo [ 29 ]	99.4	99.7	98.5	98.2	96.4	67.8
Flownomaly	99.7	99.9	99.0	98.7	96.4	67.2

Appendix

Table S5 : Full-shot anomaly detection and localization on MVTec-AD. Rotation augmentations are disabled in the full-shot setting. All values are percentages. Bold denotes the best result, and underlined denotes the second-best result.

	Image-level			Pixel-level
Method	AUROC	AUPR	$F_{1,\max}$	AUROC	AUPRO	$F_{1,\max}$
PatchCore [ 27 ]	92.8	91.4	88.0	98.5	91.1	43.7
AnomalyDINO [ 6 ]	97.6	98.0	94.5	98.8	96.1	53.8
SubspaceAD [ 18 ]	98.2	98.3	95.0	99.1	96.9	53.0
ANoCo [ 29 ]	98.7	98.9	96.1	99.2	95.3	54.7
Flownomaly	98.9	98.9	96.5	99.3	97.0	54.8

Appendix

Table S6 : Full-shot anomaly detection and localization on VisA. Rotation augmentations are disabled in the full-shot setting. All values are percentages. Bold denotes the best result, and underlined denotes the second-best result.

Figure S4 : Additional qualitative localization results on representative MVTec-AD samples under the 1-shot setting. Columns show the input image, ground-truth mask, baseline anomaly maps (AnomalyDINO [ 6 ] , SubspaceAD [ 18 ] ), and Flownomaly’s predicted anomaly map.

Figure S5 : Additional qualitative localization results on representative VisA samples under the 1-shot setting. Columns show the input image, ground-truth mask, baseline anomaly predictions (AnomalyDINO [ 6 ] , SubspaceAD [ 18 ] ), and Flownomaly.

Figure S6 : Qualitative localization results on representative Real-IAD samples under the 1-sample setting (spanning five multi-view captures of a single normal physical sample). Columns show the input query image, ground-truth mask, baseline anomaly maps (AnomalyDINO [ 6 ] , SubspaceAD [ 18 ] ), and Flownomaly.

Few-Shot Anomaly Detection

Few-shot anomaly detection uses a small set of examples to identify and localize unusual patterns in new data or domains.

See all 39 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

May 22, 2026

Beyond Normal References: Discriminative Few-Shot Anomaly Detection

This paper considers a practical few-shot anomaly detection (FSAD) setting, termed discriminative FSAD, where a limited number of both normal and anomalous examples are available as references during inference. Existing FSAD methods rely on normal-only references through normality matching, ignoring the discriminative clues in anomalous references, while directly fitting both references can overfit to the seen anomalies. We introduce IDEAL, an intrinsic deviation learning framework that leverages both reference types to learn intrinsic deviation patterns characterizing generalizable abnormality as deviations from normality. IDEAL decomposes the learning process into two novel components: 1) a Normal Variation Eraser to suppress nuisance normal variations that may lead to noisy deviations from normality, thereby highlighting anomaly-relevant deviation representations; 2) an Intrinsic Deviation Encoder to decompose these denoised deviation representations into intrinsic deviation vectors capturing the most discriminative orthogonal deviation directions. At inference, IDEAL scores query-to-normal deviations preserved after projection onto the learned intrinsic deviation vectors, enabling generalization for both seen and unseen anomalies. Extensive experiments on eight real-world datasets show that IDEAL generalizes effectively to unseen anomalies and consistently outperforms existing state-of-the-art FSAD methods. Code and data are available at https://github.com/mala-lab/IDEAL.

Huan Wang, Jun Shen, Jun Yan +1

Singapore Management University, Singapore · University of Wollongong, Australia

Few-Shot Anomaly Detection Few-Shot Learning Anomaly Detection

Ask PDF

Abstract

Few-Shot 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

Beyond Normal References: Discriminative Few-Shot Anomaly Detection