DART: Domain-Agnostic Residual Transfer for Generalist Anomaly Detection

Muhammad Aqeel; Maham Nazir; Marco Cristani; Francesco Setti

cs.CVNov 5, 2026

DART: Domain-Agnostic Residual Transfer for Generalist Anomaly Detection

Authors: Muhammad Aqeel, Maham Nazir, Marco Cristani, Francesco Setti

Organizations: University of Verona · Beihang University · Reykjavik University

Abstract

Generalist Anomaly Detection (GAD) aims to detect anomalies in unseen domains using only a few reference samples. Recent normal-abnormal guided methods improve this setting by exploiting abnormal references, but their residuals often mix transferable anomaly cues with domain-specific appearance factors such as texture, geometry, and acquisition conditions. This residual domain entanglement limits cross-domain transfer and may produce spurious activations on novel domains. We propose DART, a plug-in framework for normal-abnormal guided GAD that learns to separate residuals into anomaly-relevant and domain-specific components. DART combines a Residual Disentanglement Module, which preserves anomaly-discriminative information while suppressing domain information, with Cross-Domain Meta-Training, which forms episodes across source domains to make the separation effective. The method is architecture-agnostic and can be inserted between residual computation and the detection head of existing GAD pipelines. Experiments on MVTec AD, VisA, and BraTS show consistent improvements over multiple base methods, with the largest gains under stronger domain shifts. https://github.com/aqeeelmirza/DART

Figures & tables

Panel of Figure 1 — Figure 1: t-SNE visualization of residual features before and after DART. Colors denote defect types, while marker shapes denote source domains. (a) In the SotA baseline, residuals are fragmented and strongly influenced by domain identity: samples with the same defect type appear in multiple separated regions, while different defect types often overlap. This indicates that residuals encode both anomaly information and domain-specific factors. (b) With DART, residuals form more compact and well-separated clusters by defect type, while different domains are mixed within each cluster. This shows that DART suppresses domain-specific variation and yields residual representations that are more aligned with transferable anomaly semantics.

Figure 2: Overview of DART. (Left) Given a query $\mathbf{x}_{q}$ , normal references $\mathcal{S}_{n}$ , and an abnormal reference $\mathcal{S}_{a}$ , a frozen backbone produces features $\mathbf{f}_{q}$ , $\mathcal{F}_{n}$ , and $\mathcal{F}_{a}$ . (Center) The host residual operator computes $\mathbf{r}=\mathcal{R}(\mathbf{f}_{q};\mathcal{F}_{n},\mathcal{F}_{a})$ in residual space; the proposed Residual Disentanglement Module (RDM) decomposes $\mathbf{r}$ , and only its anomaly component $\mathbf{r}_{a}$ is passed, together with $\mathbf{f}_{q}$ , $\mathcal{F}_{n}$ , and $\mathcal{F}_{a}$ , to the unchanged host detection head. (Right) Inside RDM, the anomaly encoder $E_{a}$ (A) instantiates the transformation $T$ and extracts $\mathbf{r}_{a}$ , while the domain encoder $E_{d}$ (B) extracts the domain-specific $\mathbf{r}_{d}$ . Three losses shape the split: $\mathcal{L}_{\mathrm{suf}}$ keeps anomaly information in $\mathbf{r}_{a}$ , $\mathcal{L}_{\mathrm{adv}}$ suppresses domain information via a GRL and domain classifier $C_{d}$ , and $\mathcal{L}_{\mathrm{recon}}$ enforces that $G([\mathbf{r}_{a};\mathbf{r}_{d}])$ ( $\oplus$ ) reconstructs $\mathbf{r}$ ; $\lambda_{1},\lambda_{2}$ denote $\lambda_{\mathrm{adv}},\lambda_{\mathrm{recon}}$ . Components marked $\times$ ( $E_{d}$ , $G$ , GRL, $C_{d}$ ) are discarded at inference, leaving only $E_{a}$ active.

Target $D_{k}$	Reference $D_{i}$	Query $D_{j}$	Type
VisA–candle	MVTec–bottle	MVTec–bottle	same-domain (warmup)
VisA–candle	MVTec–bottle	MVTec–zipper	cross-domain
VisA–pcb1	MVTec–transistor	MVTec–cable	cross-domain
MVTec–bottle	VisA–candle	VisA–pcb2	cross-domain
MVTec–wood	VisA–macaroni1	VisA–fryum	cross-domain
BraTS–brain	MVTec–pill	MVTec–capsule	cross-domain

Table 1: Representative CDMT episode triplets

(D_{i},D_{j},D_{k})

. The target

D_{k}

never appears as reference or query during training.

		MVTecAD						VisA
		Image-level			Pixel-level			Image-level			Pixel-level
Setting	Method	AUROC	AP	F1-max	AUROC	PRO	F1-max	AUROC	AP	F1-max	AUROC	PRO	F1-max
$N^{1}$	SPADE [ 37 ]	81.0	90.6	90.3	91.2	83.9	42.4	79.5	82.0	80.7	95.6	84.1	35.5
	WinCLIP [ 20 ]	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 [ 21 ]	94.6	97.1	–	95.9	87.9	–	86.9	88.4	–	96.7	85.1	–
	InCTRL [ 10 ]	94.99	97.33	95.07	–	–	–	85.47	55.65	60.04	–	–	–

Table 2: Comparison on MVTecAD and VisA datasets.

N^{i}

i

normal references;

A^{1}

: one abnormal reference. +Inv : naive abnormal-reference integration without RDM. Bold / underline : best/runner-up per setting block. Shaded rows are our DART variants.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

	Host			DART Overhead (M)
Method	Params (M)	Train (H)	FPS	Train	Inference
InCTRL [ 10 ]	117.5	0.7	1.2	+3.1	+0.82
ResAD [ 11 ]	59.2	20.6	7.8	+20.1	+9.06
NAGL [ 12 ]	24.4	0.3	17.1	+4.2	+0.89

Appendix

Table 5: Efficiency comparison. DART overhead is split into training-time parameters (full RDM) and inference-time parameters (only

E_{a}

retained); DART training time is

{\approx}

1 GPU-hour per setting for all hosts.

Panel of Figure 4 — Figure 4: t-SNE visualization of token-level residuals on VisA at $k{=}2$ . (a) Without DART, each category occupies a distinct region, indicating domain texture dominates. (b) With DART, defects of the same semantic type cluster together across domains.

		MVTec AD						VisA
		Image-level			Pixel-level			Image-level			Pixel-level
Setting	Method	AUROC	AP	F1	AUROC	PRO	F1	AUROC	AP	F1	AUROC	PRO	F1
$N^{1}{+}A^{1}$	NAGL	95.54	97.63	95.40	96.53	92.66	58.03	86.23	86.54	84.59	97.20	88.76	42.41
	NAGL+DART	95.93 $\uparrow$	97.95 $\uparrow$	95.76 $\uparrow$	96.51 $\downarrow$	92.88 $\uparrow$	58.37 $\uparrow$	87.65 $\uparrow$	88.00 $\uparrow$	84.25 $\downarrow$	97.21 $\uparrow$	88.77 $\uparrow$	41.25 $\downarrow$
	InCTRL	94.99	97.33	95.07	–	–	–	85.47	55.65	60.04	–	–	–
	InCTRL+Inv	95.41	97.72	94.97	–	–	–	79.87	55.37	57.95	–	–	–

Appendix

Table 6: Full disentanglement analysis on MVTec AD and VisA. Three variants per host method: vanilla (normal only), +Inv (naive abnormal addition without RDM), +DART (with RDM).

\uparrow

\downarrow

: vs. vanilla.

Table 7: BraTS disentanglement analysis and notation reference.

Category	InCTRL						InCTRL+Inv						InCTRL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
bottle	99.92	99.98	99.21	-	-	-	99.60	99.88	98.44	-	-	-	99.68	99.90	98.44	97.68	89.63	68.43
cable	94.13	96.72	90.40	-	-	-	89.51	94.53	86.90	-	-	-	90.69	94.94	86.86	94.32	83.49	59.66
capsule	84.08	94.23	92.51	-	-	-	93.26	98.54	94.64	-	-	-	92.58	98.13	95.07	92.54	81.35	32.51
carpet	99.96	99.99	99.44	-	-	-	99.96	99.99	99.44	-	-	-	99.96	99.99	99.44	99.32	97.18	70.91
grid	98.33	99.53	98.25	-	-	-	98.75	99.63	99.12	-	-	-	99.00	99.69	99.12	97.68	91.47	35.77

Appendix

Table 8: Per-category MVTecAD results for InCTRL variants,

N^{1}+A^{1}

(1-shot) setting. Pixel-level metrics reported only for InCTRL+DART.

Category	InCTRL						InCTRL+Inv						InCTRL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
bottle	99.84	99.95	99.20	-	-	-	99.76	99.93	98.44	-	-	-	99.68	99.90	98.44	97.47	89.01	68.12
cable	94.06	96.73	90.80	-	-	-	90.37	94.96	86.32	-	-	-	91.47	95.42	87.43	94.19	82.93	59.06
capsule	81.85	93.41	92.44	-	-	-	91.70	98.26	92.64	-	-	-	92.90	98.49	94.06	92.41	82.87	31.38
carpet	99.92	99.97	99.44	-	-	-	99.96	99.99	99.44	-	-	-	99.96	99.99	99.44	99.30	96.61	69.58
grid	98.33	99.53	98.25	-	-	-	98.83	99.65	99.12	-	-	-	99.16	99.74	99.12	97.86	92.33	36.19

Appendix

Table 9: Per-category MVTecAD results for InCTRL variants,

N^{2}+A^{1}

(2-shot) setting. Pixel-level metrics reported only for InCTRL+DART.

Category	InCTRL						InCTRL+Inv						InCTRL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
bottle	100.00	100.00	100.00	-	-	-	99.84	99.95	99.21	-	-	-	99.92	99.98	99.21	97.71	89.12	68.98
cable	94.55	97.09	91.71	-	-	-	90.03	94.83	87.72	-	-	-	93.02	95.32	91.64	94.67	83.46	60.02
capsule	81.09	89.53	93.58	-	-	-	76.43	94.61	90.46	-	-	-	82.14	95.60	92.46	91.88	84.46	36.29
carpet	99.96	99.99	99.44	-	-	-	100.00	100.00	100.00	-	-	-	100.00	100.00	100.00	99.24	96.42	67.11
grid	98.83	99.65	99.12	-	-	-	98.41	99.46	96.49	-	-	-	98.83	99.61	98.21	98.19	92.88	37.96

Appendix

Table 10: Per-category MVTecAD results for InCTRL variants,

N^{4}+A^{1}

(4-shot) setting. Pixel-level metrics reported only for InCTRL+DART.

Category	InCTRL						InCTRL+Inv						InCTRL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
candle	83.40	39.46	48.57	-	-	-	80.79	65.08	65.38	-	-	-	85.60	53.64	53.97	98.61	88.59	26.32
capsules	67.76	49.66	51.61	-	-	-	67.94	49.20	51.43	-	-	-	66.45	35.86	41.86	92.84	56.32	27.16
cashew	91.00	74.09	67.86	-	-	-	89.50	78.55	72.00	-	-	-	86.63	65.56	60.00	98.38	77.92	34.23
chewinggum	97.09	93.75	88.89	-	-	-	99.23	97.00	90.48	-	-	-	96.41	90.20	86.96	98.92	79.07	50.84
fryum	95.40	89.32	85.71	-	-	-	77.10	65.86	66.67	-	-	-	94.50	86.29	82.93	92.46	72.27	30.28

Appendix

Table 11: Per-category VisA results for InCTRL variants,

N^{1}+A^{1}

(1-shot) setting. Pixel-level metrics reported only for InCTRL+DART.

Category	InCTRL						InCTRL+Inv						InCTRL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
candle	81.85	29.84	47.50	-	-	-	89.35	73.72	72.73	-	-	-	87.44	55.53	58.06	98.41	88.24	25.93
capsules	71.18	51.13	50.00	-	-	-	71.84	50.37	51.43	-	-	-	70.00	35.82	42.86	94.73	58.37	26.11
cashew	91.71	76.04	73.47	-	-	-	89.75	78.32	73.47	-	-	-	91.54	80.59	72.34	98.34	77.06	34.73
chewinggum	97.41	94.40	90.91	-	-	-	99.27	97.18	90.91	-	-	-	97.05	92.27	87.50	98.81	77.89	50.97
fryum	97.65	91.89	87.80	-	-	-	80.90	69.93	70.59	-	-	-	95.15	87.86	87.18	94.48	76.12	33.65

Appendix

Table 12: Per-category VisA results for InCTRL variants,

N^{2}+A^{1}

(2-shot) setting. Pixel-level metrics reported only for InCTRL+DART.

Category	InCTRL						InCTRL+Inv						InCTRL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
candle	78.37	23.08	43.14	-	-	-	88.98	71.93	66.67	-	-	-	88.81	56.61	54.55	98.60	88.32	25.18
capsules	73.64	50.80	48.48	-	-	-	66.23	41.34	41.38	-	-	-	73.46	48.39	48.48	94.58	58.90	26.80
cashew	90.71	78.60	73.17	-	-	-	89.38	77.17	71.70	-	-	-	90.62	80.08	71.11	98.16	77.19	35.25
chewinggum	97.09	88.34	93.33	-	-	-	93.36	90.76	86.36	-	-	-	98.00	96.79	95.45	98.81	77.97	52.46
fryum	97.80	94.26	88.89	-	-	-	89.90	81.73	76.92	-	-	-	96.10	91.63	86.49	95.95	72.60	36.45

Appendix

Table 13: Per-category VisA results for InCTRL variants,

N^{4}+A^{1}

(4-shot) setting. Pixel-level metrics reported only for InCTRL+DART.

	ResAD						ResAD+Inv						ResAD+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
bottle	100.00	100.00	100.00	98.65	96.06	76.37	99.52	99.84	99.21	98.83	95.63	78.45	100.00	100.00	100.00	98.70	96.27	77.18
cable	79.20	87.47	82.24	93.06	73.38	34.87	79.87	87.83	80.21	90.52	69.03	27.24	81.09	88.36	81.31	93.25	75.29	37.91
capsule	78.30	94.44	93.56	97.80	91.61	44.26	77.58	94.56	91.30	97.72	89.86	45.86	81.65	95.40	92.37	97.83	92.01	45.06
carpet	99.72	99.91	98.86	99.28	96.57	65.32	95.87	98.77	95.46	98.52	93.93	59.03	99.32	99.80	97.75	99.25	96.58	66.91
grid	77.94	91.28	87.69	83.14	64.08	23.34	72.77	86.23	87.02	79.00	55.30	17.13	82.21	92.99	88.00	81.35	64.97	24.04

Appendix

Table 14: Per-category MVTecAD results for ResAD variants,

N^{1}+A^{1}

(1-shot) setting.

	ResAD						ResAD+Inv						ResAD+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
bottle	99.92	99.98	99.21	98.76	96.39	77.40	99.44	99.82	99.21	98.88	95.71	78.87	100.00	100.00	100.00	98.80	96.71	78.27
cable	80.38	88.10	83.18	93.97	76.69	38.62	81.99	88.89	81.67	91.99	73.10	29.94	85.98	91.27	83.65	93.97	77.97	40.42
capsule	83.21	95.77	92.98	98.33	93.15	44.60	85.04	96.46	92.51	98.30	92.25	48.25	86.28	96.68	93.21	98.35	93.37	45.49
carpet	99.76	99.93	98.86	99.27	96.46	65.03	97.39	99.24	96.59	98.69	94.67	59.33	99.24	99.78	98.30	99.24	96.54	66.59
grid	83.46	93.17	88.71	87.53	71.09	23.31	83.29	93.40	88.00	83.28	60.24	14.73	87.80	95.42	90.91	84.75	65.16	24.39

Appendix

Table 15: Per-category MVTecAD results for ResAD variants,

N^{2}+A^{1}

(2-shot) setting.

	ResAD						ResAD+Inv						ResAD+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
bottle	99.76	99.92	99.21	98.81	96.46	77.62	99.44	99.82	99.21	98.89	95.90	78.83	100.00	100.00	100.00	98.88	96.72	78.71
cable	84.58	90.54	85.18	95.04	79.48	45.04	82.74	89.37	80.79	93.17	75.25	32.99	86.73	91.88	83.17	95.03	80.84	45.98
capsule	89.51	97.40	94.22	98.53	93.45	47.20	86.84	96.94	94.32	98.58	92.90	48.54	91.62	97.97	94.74	98.54	93.75	47.40
carpet	99.72	99.91	98.32	99.23	96.22	64.06	97.35	99.21	96.59	98.48	93.82	57.59	99.28	99.79	98.30	99.20	96.23	65.59
grid	90.48	96.48	90.00	91.17	75.78	30.55	90.48	96.18	92.44	87.91	68.39	19.93	92.40	97.08	92.56	89.29	69.78	28.21

Appendix

Table 16: Per-category MVTecAD results for ResAD variants,

N^{4}+A^{1}

(4-shot) setting.

	ResAD						ResAD+Inv						ResAD+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
candle	86.25	84.83	80.24	98.26	93.44	32.97	85.09	84.29	80.67	98.20	93.76	27.06	89.19	88.35	84.00	98.96	95.14	29.39
capsules	70.74	78.74	78.75	95.32	76.09	29.60	58.38	73.05	76.92	95.33	60.69	28.89	58.62	68.84	76.98	94.13	71.42	18.78
cashew	91.08	95.16	87.53	96.22	89.72	41.13	93.18	96.86	90.45	95.85	89.31	38.56	93.12	96.68	91.79	97.53	90.83	45.43
chewinggum	97.48	97.88	95.22	98.49	82.33	64.47	95.56	98.10	93.94	98.38	75.58	55.96	98.30	99.27	96.41	98.96	82.17	54.46
fryum	77.49	88.50	81.64	90.67	70.89	35.00	70.34	83.46	81.15	90.31	67.66	31.89	76.78	88.38	82.01	90.68	70.72	31.02

Appendix

Table 17: Per-category VisA results for ResAD variants,

N^{1}+A^{1}

(1-shot) setting.

	ResAD						ResAD+Inv						ResAD+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
candle	84.83	82.11	81.59	98.35	93.35	31.24	83.63	82.18	78.22	98.23	93.59	25.66	86.53	86.02	81.38	98.83	94.93	27.32
capsules	73.19	82.16	79.27	97.18	80.27	49.36	67.37	77.79	77.22	96.24	62.53	39.99	66.88	73.55	79.34	97.18	76.53	36.21
cashew	91.83	95.37	88.25	96.48	90.40	42.10	94.08	97.47	91.01	96.47	90.91	41.06	95.18	97.85	93.14	97.76	91.65	47.64
chewinggum	97.53	97.81	95.54	98.62	81.91	64.10	98.36	99.24	95.10	98.44	76.59	57.14	98.68	99.40	96.52	98.92	82.60	53.90
fryum	82.05	90.31	83.23	90.92	71.48	36.17	68.66	84.15	81.33	90.65	70.59	33.21	72.38	85.81	80.69	91.11	71.99	33.82

Appendix

Table 18: Per-category VisA results for ResAD variants,

N^{2}+A^{1}

(2-shot) setting.

	ResAD						ResAD+Inv						ResAD+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
candle	86.52	83.89	82.31	98.55	94.18	31.39	84.32	81.67	80.00	98.54	94.25	27.13	88.63	86.94	81.82	98.92	95.25	27.99
capsules	79.18	86.78	79.96	97.56	82.94	52.76	68.82	80.26	77.65	96.88	67.59	44.73	73.55	81.87	81.55	97.45	81.53	45.19
cashew	92.92	96.33	89.69	96.73	91.22	42.94	94.66	97.56	90.32	96.65	91.37	41.93	95.48	97.99	92.86	97.88	91.54	48.84
chewinggum	98.22	98.59	97.37	98.73	82.11	63.76	98.52	99.33	96.91	98.66	76.40	59.10	99.30	99.70	98.49	98.94	81.02	54.25
fryum	87.55	93.85	84.88	91.56	73.35	38.63	78.02	88.70	83.33	91.40	72.04	37.50	83.02	91.90	83.65	91.81	73.40	38.30

Appendix

Table 19: Per-category VisA results for ResAD variants,

N^{4}+A^{1}

(4-shot) setting.

	NAGL						NAGL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
bottle	98.78	99.52	98.70	98.31	95.61	76.32	98.89	99.57	98.95	98.20	95.58	76.16
cable	90.09	94.35	87.96	95.48	88.10	58.16	92.62	95.89	90.59	95.32	88.16	57.92
capsule	90.68	97.91	94.17	97.95	96.36	45.57	91.18	98.02	94.77	98.00	96.48	46.31
carpet	100.00	100.00	100.00	99.29	98.40	68.15	100.00	100.00	100.00	99.36	98.52	68.89
grid	99.81	99.93	99.41	99.23	96.68	48.29	99.75	99.91	99.12	99.29	96.92	49.61

Appendix

Table 20: Per-category MVTecAD results for NAGL variants,

N^{1}+A^{1}

(1-shot) setting.

	NAGL						NAGL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
bottle	98.99	99.57	99.48	98.22	95.81	76.47	98.99	99.58	99.48	98.17	95.72	76.36
cable	92.73	95.86	89.74	95.72	88.70	58.87	92.69	95.91	89.78	95.54	88.40	58.76
capsule	90.79	97.89	94.81	98.01	96.60	46.37	90.63	97.86	94.67	98.06	96.66	47.01
carpet	100.00	100.00	100.00	99.33	98.46	68.59	100.00	100.00	100.00	99.35	98.48	68.75
grid	99.81	99.93	99.42	99.31	97.00	49.44	99.83	99.94	99.42	99.32	97.02	49.57

Appendix

Table 21: Per-category MVTecAD results for NAGL variants,

N^{2}+A^{1}

(2-shot) setting.

	NAGL						NAGL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
bottle	99.74	99.91	99.21	98.27	95.77	76.65	99.79	99.93	99.21	98.17	95.61	76.37
cable	93.93	96.63	90.58	95.86	89.22	60.03	95.01	97.26	91.86	95.82	89.35	60.35
capsule	96.04	99.19	95.87	98.26	97.05	51.79	96.48	99.29	96.04	98.28	97.08	52.18
carpet	100.00	100.00	100.00	99.33	98.43	68.25	100.00	100.00	100.00	99.35	98.49	68.63
grid	99.89	99.96	99.71	99.33	97.06	49.99	99.83	99.94	99.42	99.33	97.07	50.14

Appendix

Table 22: Per-category MVTecAD results for NAGL variants,

N^{4}+A^{1}

(4-shot) setting.

	NAGL						NAGL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
candle	92.99	92.97	86.12	98.68	93.40	40.61	91.41	91.76	84.16	98.83	93.41	39.48
capsules	97.80	98.54	95.19	97.73	81.69	48.64	95.67	97.30	93.20	97.16	80.92	36.30
cashew	90.84	94.84	89.81	98.77	94.69	54.18	86.08	93.42	85.19	98.88	95.25	57.43
chewinggum	97.48	98.94	95.29	99.26	94.98	69.02	96.68	98.69	95.92	99.03	93.88	68.96
fryum	95.96	98.21	93.53	96.48	83.27	45.74	94.12	97.23	90.64	95.40	82.05	39.33

Appendix

Table 23: Per-category VisA results for NAGL variants,

N^{1}+A^{1}

(1-shot) setting.

	NAGL						NAGL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
candle	93.94	94.15	87.32	98.88	93.92	42.41	92.37	93.15	85.02	98.70	92.42	37.88
capsules	98.20	98.84	95.19	97.38	80.61	48.09	95.67	97.29	92.52	97.27	80.81	48.33
cashew	86.84	93.72	86.12	98.58	93.56	56.20	93.78	97.10	92.08	99.06	96.76	59.06
chewinggum	96.94	98.76	95.29	99.16	94.52	69.94	98.20	99.20	95.43	99.16	94.80	68.58
fryum	96.50	98.40	94.36	96.10	82.38	44.64	96.88	98.62	95.05	95.88	82.63	43.97

Appendix

Table 24: Per-category VisA results for NAGL variants,

N^{2}+A^{1}

(2-shot) setting.

	NAGL						NAGL+DART
Category	I-AUC	AP	I-F1	P-AUC	PRO	P-F1	I-AUC	AP	I-F1	P-AUC	PRO	P-F1
candle	92.75	93.07	87.08	98.84	93.62	42.21	92.50	93.27	85.31	98.81	93.10	39.42
capsules	98.22	98.94	94.47	97.39	80.21	51.93	95.28	97.00	92.52	98.02	83.05	49.48
cashew	92.26	96.57	87.91	98.63	93.76	58.16	91.92	96.27	89.45	98.89	96.90	57.77
chewinggum	97.28	98.89	95.88	99.17	94.67	69.84	97.34	98.90	94.95	99.12	94.51	67.82
fryum	96.90	98.59	94.85	96.11	82.62	45.26	97.32	98.79	95.43	95.94	83.50	44.18

Appendix

Table 25: Per-category VisA results for NAGL variants,

N^{4}+A^{1}

(4-shot) setting.

Anomaly Detection

Anomaly detection finds rare or unexpected patterns in data, from defective products and system faults to fraud and security threats.

See all 689 papers

Apr 6, 2026

InCTRLv2: Generalist Residual Models for Few-Shot Anomaly Detection and Segmentation

While recent anomaly detection (AD) methods have made substantial progress in recognizing abnormal patterns within specific domains, most of them are specialist models that are trained on large training samples from a specific target dataset, struggling to generalize to unseen datasets. To address this limitation, the paradigm of Generalist Anomaly Detection (GAD) has emerged in recent years, aiming to learn a single generalist model to detect anomalies across diverse domains without retraining. To this end, this work introduces InCTRLv2, a novel few-shot Generalist Anomaly Detection and Segmentation (GADS) framework that significantly extends our previously proposed GAD model, InCTRL. Building on the idea of learning in-context residuals with few-shot normal examples to detect anomalies as in InCTRL, InCTRLv2 introduces two new, complementary perspectives of anomaly perception under a dual-branch framework. This is accomplished by two novel modules upon InCTRL: i) Discriminative Anomaly Score Learning (DASL) with both normal and abnormal data in the main branch, which learns a semantic-guided abnormality and normality space that supports the classification of query samples from both the abnormality and normality perspectives; and ii) One-class Anomaly Score Learning (OASL) using only the normal data, which learns generalized normality patterns in a semantic space via an auxiliary branch, focusing on detecting anomalies through the lens of normality solely. Both branches are guided by rich visual-text semantic priors encoded by large-scale vision-language models. Together, they offer a dual semantic perspective for AD: one emphasizes normal-abnormal discriminations, while the other emphasizes normality-deviated semantics. Extensive experiments on ten AD datasets demonstrate that InCTRLv2 achieves SotA performance in both anomaly detection and segmentation tasks across various settings.

Jiawen Zhu, Mengjia Niu, Guansong Pang

School of Computing and Information Systems, Singapore Management University, Singapore. · Dyson School of Design Engineering, Imperial College London, United Kingdom.

Anomaly Detection Few-Shot Anomaly Detection

Ask PDF

Mar 10, 2026

TA-GGAD: Testing-time Adaptive Graph Model for Generalist Graph Anomaly Detection

A significant number of anomalous nodes in the real world, such as fake news, noncompliant users, malicious transactions, and malicious posts, severely compromises the health of the graph data ecosystem and urgently requires effective identification and processing. With anomalies that span multiple data domains yet exhibit vast differences in features, cross-domain detection models face severe domain shift issues, which limit their generalizability across all domains. This study identifies and quantitatively analyzes a specific feature mismatch pattern exhibited by domain shift in graph anomaly detection, which we define as the Anomaly Disassortativity issue (

\mathcal{AD}

). Based on the modeling of the issue

\mathcal{AD}

, we introduce a novel graph foundation model for anomaly detection. It achieves cross-domain generalization in different graphs, requiring only a single training phase to perform effectively across diverse domains. The experimental findings, based on fourteen diverse real-world graphs, confirm a breakthrough in the model's cross-domain adaptation, achieving a pioneering state-of-the-art (SOTA) level in terms of detection accuracy. In summary, the proposed theory of

\mathcal{AD}

provides a novel theoretical perspective and a practical route for future research in generalist graph anomaly detection (GGAD). The code is available at https://anonymous.4open.science/r/Anonymization-TA-GGAD/.

Xiong Zhang, Hong Peng, Changlong Fu +3

School of Software and AI, Yunnan University, Kunming, China

Anomaly Detection Cross-Dataset Generalization Graph Foundation Models Test-Time Adaptation

Ask PDF

May 12, 2026

Anomaly-Aware Vision-Language Adapters for Zero-Shot Anomaly Detection

Zero-shot anomaly detection aims to identify defects in unseen categories without target-specific training. Existing methods usually apply the same feature transformation to all samples, treating normal and anomalous data uniformly despite their fundamentally asymmetric distributions, compact normals versus diverse anomalies. We instead exploit this natural asymmetry by proposing AVA-DINO, an anomaly-aware vision-language adaptation framework with dual specialized branches for normal and anomalous patterns that adapt frozen DINOv3 visual features. During training on auxiliary data, the two branches are learned jointly with a text-guided routing mechanism and explicit routing regularization that encourages branch specialization. At test time, only the input image and fixed, predefined language descriptions are used to dynamically combine the two branches, enabling an asymmetric activation. This design prevents degenerate uniform routing and allows context-specific feature transformations. Experiments across nine industrial and medical benchmarks demonstrate state-of-the-art performance, achieving 93.5% image-AUROC on MVTec-AD and strong cross-domain generalization to medical imaging without domain-specific fine-tuning. https://github.com/aqeeelmirza/AVA-DINO

Muhammad Aqeel, Maham Nazir, Uzair Khan +2

Dept. of Engineering for Innovation Medicine, University of Verona, Italy · School of Computer Science and Engineering, Beihang University, China · Dept. of Engineering for Innovation +1

VLM Adaptation Medical Image Anomaly Detection Industrial Anomaly Detection Vision Foundation Model Adaptation Multimodal Anomaly Detection Zero-Shot Anomaly Detection

Ask PDF

Abstract

Anomaly Detection

Explore similar work

InCTRLv2: Generalist Residual Models for Few-Shot Anomaly Detection and Segmentation

TA-GGAD: Testing-time Adaptive Graph Model for Generalist Graph Anomaly Detection

Anomaly-Aware Vision-Language Adapters for Zero-Shot Anomaly Detection