Co-Evolutionary Prompt Optimization with Cross-Category Transfer for Zero-Shot Anomaly Detection

Sisi Zhu; Changwei Yu; Renshuai Tao; Zhenliang Ni

cs.CVAug 29, 2026

Co-Evolutionary Prompt Optimization with Cross-Category Transfer for Zero-Shot Anomaly Detection

Authors: Sisi Zhu, Changwei Yu, Renshuai Tao, Zhenliang Ni

Organizations: Institute of Information Science, Beijing Jiaotong University, Beijing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China

Abstract

Zero-shot anomaly detection (ZSAD) has gained significant attention for its practical value in industrial inspection. Recently, CLIP-based approaches have been widely adopted in ZSAD due to their strong vision-language generalization capabilities. However, existing methods commonly employ continuous prompt embeddings for prompt optimization and encode semantics in latent vectors, which lack interpretability and scalability. To this end, we propose CoEvoAD, a co-evolutionary framework for discrete prompt selection. CoEvoAD performs prompt search in the discrete natural-language space using an evolutionary algorithm. Candidate prompts are iteratively generated, evaluated, and selected throughout population evolution, thus preserving the interpretability and composability of natural language. Furthermore, we introduce a Cross-Category Transfer Objective (CCTO), which treats held-out source categories as proxies for unseen categories and scores prompt rules based on their estimated cross-category transferability, effectively improving cross-category generalization. Extensive experiments are conducted to validate the effectiveness of CoEvoAD, and the results show that it achieves state-of-the-art performance across multiple anomaly detection datasets. The code is available at https://github.com/rstao-bjtu/CoEvoAD.

Figures & tables

Figure 1: Prompt regimes for ZSAD: handcrafted templates, continuous prompts, and CoEvoAD.

Figure 2: Overview of CoEvoAD. A frozen source-trained scorer evaluates role-separated normal and abnormal rule populations; CCTO uses held-out source categories to select rules for unseen-category inference.

Method $\rightarrow$ Dataset $\downarrow$	WinCLIP (CVPR’23)		AnomalyCLIP (ICLR’24)		AdaCLIP (ECCV’24)		Bayes-PFL (CVPR’25)		MRAD (ICLR’26)		CoEvoAD (Ours)
	AUROC $\uparrow$	AP $\uparrow$	AUROC $\uparrow$	AP $\uparrow$	AUROC $\uparrow$	AP $\uparrow$	AUROC $\uparrow$	AP $\uparrow$	AUROC $\uparrow$	AP $\uparrow$	AUROC $\uparrow$	AP $\uparrow$
MVTec-AD	91.8	95.1	91.5	96.2	92.0	96.4	92.3	96.7	94.0	97.4	93.4	96.8
VisA	78.1	77.5	82.1	85.4	83.0	84.9	87.0	89.2	85.7	88.3	87.4	89.7
BTAD	83.3	84.1	89.1	91.1	91.6	92.4	93.2	96.5	92.4	94.2	94.4	96.1
KSDD2	93.5	77.9	92.1	77.8	95.9	95.9	97.3	97.9	95.1	88.9	97.4	97.9
DAGM	89.6	90.4	95.6	94.6	96.5	95.7	97.7	97.0	98.4	98.6	98.2	97.6

Table 1: Image-level benchmark comparison (%) across six anomaly detection datasets. Bold/underline denote best/second-best; “–” marks entries not reported by the original paper.

Method $\rightarrow$ Dataset $\downarrow$	WinCLIP (CVPR’23)		AnomalyCLIP (ICLR’24)		AdaCLIP (ECCV’24)		Bayes-PFL (CVPR’25)		MRAD (ICLR’26)		CoEvoAD (Ours)
	AUROC $\uparrow$	PRO $\uparrow$	AUROC $\uparrow$	PRO $\uparrow$	AUROC $\uparrow$	PRO $\uparrow$	AUROC $\uparrow$	PRO $\uparrow$	AUROC $\uparrow$	PRO $\uparrow$	AUROC $\uparrow$	PRO $\uparrow$
MVTec-AD	85.1	64.6	91.1	81.4	86.8	33.8	91.8	87.4	93.0	86.8	92.2	87.9
VisA	79.6	56.8	95.5	87.0	95.1	71.3	95.6	88.9	95.9	88.0	95.8	89.4
BTAD	71.4	32.8	93.3	69.3	87.7	17.1	93.9	76.6	95.4	72.8	94.7	81.9
KSDD2	97.9	91.2	99.1	85.6	99.4	92.7	96.1	70.8	98.9	95.6	99.6	98.5
DAGM	83.2	55.4	99.1	93.6	97.0	40.9	99.3	98.0	97.4	90.3	99.5	98.3

Table 2: Pixel-level benchmark comparison (%; AUROC and Per-Region Overlap (PRO)). Bold/underline denote best/second-best across the six datasets; “–” marks entries not reported by the original paper.

Target	Metric	Control	CoEvoAD	$\Delta$
MVTec-AD	Image AUROC $\uparrow$	93.01	93.39	+0.38
	Pixel AP $\uparrow$	47.48	47.72	+0.24
	Pixel F1 $\uparrow$	47.69	47.96	+0.27
VisA	Image AUROC $\uparrow$	87.05	87.43	+0.38
	Pixel AP $\uparrow$	30.51	31.11	+0.60
	Pixel F1 $\uparrow$	36.21	36.72	+0.51

Table 3: Matched comparison. Both systems use the same frozen source-trained scorer and protocol; the control uses class-name prompts without rule search.

\Delta=\mathrm{CoEvoAD}-\mathrm{Control}

. Seed 111; three-seed means and standard deviations for the two primary directions are reported in Table 17 .

Figure 3: Per-image anomaly score distributions on three representative MVTec-AD categories (Pill, Toothbrush, Hazelnut) under the VisA

\to

MVTec-AD matched protocol. Rows compare AnomalyCLIP, Bayes-PFL, and CoEvoAD; blue/red denote normal/abnormal samples; per-cell AUROC is shown.

Dir.	Metric	Mean $\Delta$	Bot- $k$ $\Delta$	Worst $\Delta$
MVTec	Pixel AP $\uparrow$	+0.24	+0.51	$-$ 0.48
MVTec	Pixel F1 $\uparrow$	+0.27	+0.45	$-$ 0.41
VisA	Pixel AP $\uparrow$	+0.60	+0.78	$-$ 3.27
VisA	Pixel F1 $\uparrow$	+0.51	+0.77	$-$ 3.37

Table 4: Per-category

\Delta

statistics (mean, bottom-

k

, worst) vs the matched control.

Figure 4: Qualitative localization examples on MVTec-AD and VisA. Columns show representative target categories; rows compare ground-truth contours and anomaly maps from CLIP-based baselines and CoEvoAD. Green contours denote annotated anomaly regions; warmer colors indicate higher anomaly scores.

Category	Normal rule	Abnormal rule
screw	clean surface screw	broken shape with missing parts
bottle	normal bottle	irregular visible surface bottle
grid	regular standard grid	faulty flawed abnormal grid
toothbrush	a photo of a healthy toothbrush	faulty toothbrush
pcb2	flawless fine pcb2	structural defect like a hole or cut
cashew	a photo of healthy cashew	faulty flawed cashew

Table 6: Representative natural-language rules selected by CoEvoAD (the leading learned soft-context token is omitted for compactness).

Target	Metric	Baseline	CoEvo	CoEvo +CCTO
MVTec-AD	Image AUROC $\uparrow$	93.01	93.06	93.39
	Pixel AP $\uparrow$	47.48	47.52	47.72
	Pixel F1 $\uparrow$	47.69	47.67	47.96
VisA	Image AUROC $\uparrow$	87.05	87.02	87.43
	Pixel AP $\uparrow$	30.51	30.98	31.11
	Pixel F1 $\uparrow$	36.21	36.33	36.72

Table 7: Component ablation comparing matched class-name prompt control, role-separated co-evolution (CoEvo), and CoEvoAD (CoEvo

+

CCTO). Bold marks the best score per row. Seed 111 ( Table 17 reports multi-seed variation).

Method	Backbone / res.	ms/img
Bayes-PFL (reproduced)	ViT-L/14@336px / 518	221.0
AdaCLIP (released ckpt)	ViT-L/14@336px / 518	114.56
MRAD (released ckpt)	ViT-L/14@336px / 518	103.55
AnomalyCLIP (reproduced)	ViT-L/14@336px / 518	75.97
WinCLIP † (training-free)	ViT-B/16+240 / 240	56.08
Class-name control	ViT-L/14@336px / 518	55.00

Table 8: Per-image inference latency (ms, mean over 300 timed images), timed on the VisA

\to

MVTec-AD direction under one protocol.

\dagger

: WinCLIP runs its native backbone and resolution, disclosed for completeness but not directly comparable.

Appendix figures & tables45 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Classes	Normal	Anomalous
MVTec-AD Bergmann et al. (2019)	15	467	1,258
VisA Zou et al. (2022)	12	962	1,200
BTAD Mishra et al. (2021)	3	451	290
KSDD2 Božič et al. (2021)	1	356	356
RSDD Niu et al. (2021)	1	387	387
DAGM (7-class subset) Wieler et al. (2007)	7	1,350	1,350

Appendix

Table 9: Test-set statistics of the evaluation benchmarks. KSDD2 and RSDD are rebuilt from their official splits into balanced test sets following Bayes-PFL Qu et al. (2025) ; the DAGM row refers to the 7-class subset used in our protocol ( Table 20 ).

Component	Setting
Backbone	CLIP ViT-L/14@336px
Input resolution	$518\times 518$
Stage 1 epochs / optimiser	30 / AdamW, cosine annealing
Stage 1 learning rates	prompt $1{\times}10^{-3}$ , other $1{\times}10^{-4}$
Prompt groups $M$ / context $L_{c}$ / state $L_{s}$	3 / 5 / 5
Search population $N$	16

Appendix

Table 10: Full implementation and hyperparameters.

Variant	AUROC	AP px
Full training setup (control)	93.01	47.48
$-$ mask-guided crop	92.94 ( $-$ 0.07)	46.35 ( $-$ 1.13)
$-$ inter-role margin	92.83 ( $-$ 0.18)	47.04 ( $-$ 0.44)
$-$ agnostic reg.	93.01 ( $\pm$ 0.00)	47.09 ( $-$ 0.39)

Appendix

Table 11: Ablation of Stage 1 training components on the class-name control. Each row removes one component from the fixed scorer training.

Dir.	$\|\mathcal{C}_{\mathrm{src}}\|$	$N$	$G$	$K$	Evals/role	Wall-clock
MVTec (src VisA)	12	16	5	3	960	$\sim$ 3 h 47 m
VisA (src MVTec-AD)	15	16	5	3	1200	$\sim$ 5 h 34 m

Appendix

Table 12: Rule-search cost.

|\mathcal{C}_{\mathrm{src}}|

: number of source categories. Evals/role

=N{\times}G{\times}|\mathcal{C}_{\mathrm{src}}|

; both roles are scored each generation.

$\|\mathcal{C}_{\mathrm{src}}\|$	Evals/role	Wall-clock (h)	s/eval
3	240	0.34	5.1
6	480	1.12	8.4
9	720	2.38	11.9
12	960	3.71	13.9

Appendix

Table 13: Measured search-cost scaling with the source-category count (VisA

\to

MVTec-AD configuration, single RTX 4090). Evals/role

=N{\times}G{\times}|\mathcal{C}_{\mathrm{src}}|

; s/eval is wall-clock divided by this per-role count.

			MVTec-AD		VisA
Method	Role-sep.	CCTO	AUROC img	AP px	AUROC img	AP px
Class-name control (R0)	✗	✗	93.01	47.48	87.05	30.51
Random prompt-pair search	✗	✗	93.09	47.47	86.10	31.03
Random prompt-pair search $+$ CCTO	✗	✓	93.23	47.45	86.90	30.80
Single-population evolutionary algorithm	✗	✗	93.15	47.43	86.81	30.58
Single-population evolutionary algorithm $+$ CCTO	✗	✓	93.24	47.54	86.90	30.62

Appendix

Table 14: Extended same-budget prompt-search control table with CCTO-augmented variants. All search rows use the same frozen scorer and candidate budget; R0 is the non-search class-name reference; seed 111.

	AUROC img			AP px			F1 px
Category	C	E	$\Delta$	C	E	$\Delta$	C	E	$\Delta$
bottle	94.76	94.84	+0.08	67.19	66.97	$-$ 0.22	63.09	62.92	$-$ 0.17
cable	82.27	83.36	+1.09	12.49	12.01	$-$ 0.48	20.09	19.98	$-$ 0.11
capsule	94.50	94.30	$-$ 0.20	36.72	36.67	$-$ 0.05	39.85	39.44	$-$ 0.41
carpet	100.00	100.00	+0.00	82.69	82.98	+0.29	74.93	75.22	+0.29
grid	100.00	100.00	+0.00	41.38	40.97	$-$ 0.41	42.70	42.75	+0.05

Appendix

Table 15: Per-category breakdown for the MVTec direction (VisA

\to

MVTec-AD). C: strict class-name control. E: CoEvoAD.

\Delta=

-

C (pp). Image AUROC / pixel AP / pixel F1. Values are single-seed (seed 111); 3-seed means per routed category are reported in Table 24 , and per-category three-seed mean

\pm

std in Table 18 .

	AUROC img			AP px			F1 px
Category	C	E	$\Delta$	C	E	$\Delta$	C	E	$\Delta$
candle	90.02	90.67	+0.65	33.69	33.53	$-$ 0.16	44.08	43.55	$-$ 0.53
capsules	92.95	91.95	$-$ 1.00	49.56	52.27	+2.71	54.92	55.78	+0.86
cashew	93.84	93.36	$-$ 0.48	36.35	37.40	+1.05	42.16	42.68	+0.52
chewinggum	97.38	97.00	$-$ 0.38	83.18	83.12	$-$ 0.06	75.45	75.53	+0.08
fryum	90.68	90.92	+0.24	27.95	27.60	$-$ 0.35	33.19	32.92	$-$ 0.27

Appendix

Table 16: Per-category breakdown for the MVTec-AD

\to

VisA direction. C: strict class-name control. E: CoEvoAD (MVTec-AD

\to

VisA main config).

\Delta=

-

C (pp). The MVTec-AD

\to

VisA gain is dominated by a few categories ( pipe_fryum +4.05 AP px , pcb4

-

3.27 AP px ). Values are single-seed (seed 111); 3-seed means per routed category are reported in Table 25 ( pipe_fryum

+1.96

, pcb4

-2.18

AP px ), and per-category three-seed mean

\pm

std in Table 18 .

	VisA $\to$ MVTec-AD		MVTec-AD $\to$ VisA
	AUROC img	AP px	AUROC img	AP px
Seed 111	93.39	47.72	87.43	31.11
Seed 222	93.13	47.37	87.05	30.79
Seed 333	93.26	47.45	87.46	31.19
Mean $\pm$ std	93.26 $\pm$ 0.13	47.51 $\pm$ 0.18	87.31 $\pm$ 0.23	31.03 $\pm$ 0.21

Appendix

Table 17: Multi-seed stability over seeds 111, 222, and 333. We report per-seed results together with mean and standard deviation under the matched protocol.

VisA $\to$ MVTec-AD
Category	AUROC img	AP px
bottle	94.84 $\pm$ 0.00	67.03 $\pm$ 0.20
cable	83.55 $\pm$ 0.32	12.09 $\pm$ 0.20
capsule	94.31 $\pm$ 0.02	36.60 $\pm$ 0.06
carpet	100.00 $\pm$ 0.00	82.77 $\pm$ 0.19
grid	100.00 $\pm$ 0.00	41.16 $\pm$ 0.17

Appendix

Table 18: Per-category three-seed mean

\pm

std of CoEvoAD over the same three independent search runs as Table 17 (deployed configuration per direction). Std is the sample standard deviation over seeds 111/222/333.

Source	Target	AUROC img C $\to$ E	$\Delta$ AUROC px	$\Delta$ AUPRO	$\Delta$ AP px	$\Delta$ F1 px
MVTec	BTAD	92.81 $\to$ 92.68	+0.45	+0.47	+0.61	+1.56
MVTec	DAGM	93.82 $\to$ 93.56	$-$ 0.12	$-$ 0.46	+0.81	+0.35
MVTec	DTD	93.77 $\to$ 93.99	+0.12	+0.25	+0.04	+0.32
MVTec	KSDD2	96.04 $\to$ 95.94	+0.01	+0.09	+0.14	+0.21
MVTec	RSDD	97.69 $\to$ 97.62	+0.02	+0.10	+2.91	+0.99
MVTec	Mean	94.83 $\to$ 94.76	+0.10	+0.09	+0.90	+0.69

Appendix

Table 19: External industrial benchmarks, all five targets, both source conventions. AUROC img : control

\to

CoEvoAD (%).

\Delta

columns: CoEvoAD

-

matched strict control (pp).

Same 7-class DAGM subset
Method	DAGM	DTD-Synthetic
AnomalyCLIP †	96.3	94.3
Bayes-PFL †	97.35	$\mathbf{96.28}$
CoEvoAD	$\mathbf{98.19}$	94.25
Original publication reports
WinCLIP	87.6	83.9

Appendix

Table 20: Image-level AUROC (%) on texture-oriented external targets.

\dagger

denotes evaluation on our 7-class DAGM subset; unmarked DAGM numbers follow the standard 10-class reporting convention.

Category	C (Bayes-PFL)	E (CoEvoAD)	$\Delta$
blotchy	94.06	75.75	$-18.31$
matted1	88.10	80.95	$-7.15$
marbled2	99.00	96.50	$-2.50$
fibrous	99.94	98.69	$-1.25$
mesh	91.15	92.64	$+1.49$
perforated	93.00	96.44	$+3.44$

Appendix

Table 21: Per-class image-level AUROC (%) on DTD-Synthetic. C: Bayes-PFL (matched VisA-source reproduction). E: CoEvoAD.

\Delta=

-

Dir.	Metric	Mean $\Delta$	95% CI	#Up	#Down	#Tie
MVTec	AUROC img	+0.38	[+0.10, +0.66]	8	2	5
	AP px	+0.24	[ $-$ 0.03, +0.51]	9	6	0
	F1 px	+0.27	[+0.06, +0.47]	11	4	0
VisA	AUROC img	+0.37	[ $-$ 0.19, +0.93]	7	5	0
	AP px	+0.60	[ $-$ 0.38, +1.58]	7	5	0
	F1 px	+0.51	[ $-$ 0.47, +1.49]	8	4	0

Appendix

Table 22: Paired category-level mean differences with 95% confidence intervals.

\Delta=

CoEvoAD

-

matched control, averaged over categories (

n{=}15

for the VisA

\to

MVTec-AD direction,

n{=}12

for the MVTec-AD

\to

VisA direction). CI is the normal-approximation 95% interval of the per-category

\Delta

. Differences of

\pm

0.01 from Table 3 reflect different rounding paths (aggregate vs per-category).

Route	Sem	Tmpl	AUROC img	AP px	AUROC img	AP px
			MVTec-AD		VisA
R0 (control)	✗	✗	93.01	47.48	87.05	30.51
R1 (sem-only)	✓	✗	93.39	47.72	87.21	30.85
R2 (tmpl-only)	✗	✓	93.13	47.49	87.11	31.51
R3 (both)	✓	✓	93.39	47.72	87.21	30.85

Appendix

Table 23: Test-time transfer ablation. R0 is the class-name control; Sem/Tmpl denote nearest-source/template transfer.

Category	Donor	Sim	$\Delta$ AUROC img	$\Delta$ AP px	$\Delta$ F1 px
bottle	candle	0.810	+0.08	$-$ 0.16	$-$ 0.09
cable	macaroni1	0.758	+1.28	$-$ 0.40	+0.01
capsule	capsules	0.903	$-$ 0.19	$-$ 0.12	$-$ 0.34
carpet	fryum	0.799	+0.00	+0.08	+0.07
grid	fryum	0.803	+0.00	$-$ 0.22	+0.11
hazelnut	cashew	0.848	+0.02	+0.90	+0.56

Appendix

Table 24: Per-class routing audit, VisA

\to

MVTec-AD: routed donor, semantic similarity, and deltas vs. the class-name control (pp, 3-seed means; single-seed values in Table 15 , policy tags in Table 31 ). The control is a single run; seed variance is method-side (cf. Table 17 ).

Category	Donor	Sim	$\Delta$ AUROC img	$\Delta$ AP px	$\Delta$ F1 px
candle	bottle	0.810	+0.83	+0.75	$-$ 0.06
capsules	capsule	0.903	$-$ 0.65	+2.95	+1.66
cashew	hazelnut	0.848	$-$ 0.34	+0.04	$-$ 0.15
chewinggum	toothbrush	0.716	$-$ 0.23	+0.03	+0.12
fryum	pill	0.819	+0.21	+0.71	+0.39
macaroni1	pill	0.826	$-$ 0.39	+0.07	+0.88

Appendix

Table 25: Per-class routing audit, MVTec-AD

\to

VisA (pp, 3-seed means; single-seed values in Table 16 , policy tags in Table 32 ).

	MVTec	VisA
Forced route	AUROC img / AP px / F1 px
Actual route (anchor, absolute)	93.39 / 47.72 / 47.96	87.43 / 31.11 / 36.72
Random donor (expectation)	$-$ 0.21 / $-$ 0.25 / $-$ 0.28	$-$ 0.25 / $-$ 0.11 / $-$ 0.02
Worst donor (adversarial)	$-$ 0.54 / $-$ 0.95 / $-$ 0.88	$-$ 2.26 / $-$ 2.47 / $-$ 2.35
Oracle donor (not zero-shot)	+0.14 / +0.48 / +0.35	+1.47 / +2.29 / +2.29

Appendix

Table 26: Donor-forced counterfactual aggregates (

\Delta

vs. the actual route, pp; negative = worse). Anchor row: absolute scores of the actual route under this surface. Oracle uses target labels (not zero-shot; reference only).

	AUROC img				AP px				F1 px
Category	Act.	Rand.	Worst	Orac.	Act.	Rand.	Worst	Orac.	Act.	Rand.	Worst	Orac.
bottle	94.84	94.92	94.76	95.16	66.97	67.88	66.79	68.46	62.92	63.62	62.78	64.09
cable	83.36	83.09	82.38	83.85	12.01	12.61	12.01	13.19	19.98	20.36	19.98	20.99
capsule	94.30	94.46	94.14	94.81	36.67	36.89	36.36	37.36	39.44	39.84	39.34	40.40
carpet	100.00	100.00	100.00	100.00	82.98	82.76	82.50	83.03	75.22	74.98	74.76	75.22
grid	100.00	99.99	99.92	100.00	40.97	41.22	40.80	41.63	42.75	42.70	42.11	43.04

Appendix

Table 27: Per-class donor-forced counterfactual envelope, VisA

\to

MVTec-AD (absolute scores under the forced-route surface; in-family comparison only). Act.: actual route; Rand.: uniform-random expectation over donors; Worst/Orac.: per-class minimum/maximum over the donor bank (oracle uses target labels; not zero-shot).

	AUROC img				AP px				F1 px
Category	Act.	Rand.	Worst	Orac.	Act.	Rand.	Worst	Orac.	Act.	Rand.	Worst	Orac.
candle	90.67	90.63	89.79	91.75	33.53	33.96	31.99	36.12	43.55	43.94	42.61	45.79
capsules	91.95	92.81	91.60	94.03	52.27	51.45	44.31	55.32	55.78	56.38	49.35	59.93
cashew	93.36	93.33	92.46	94.20	37.40	37.41	35.47	40.90	42.68	43.01	41.50	45.55
chewinggum	97.00	97.15	96.88	97.48	83.12	82.99	81.19	83.82	75.53	75.43	74.35	76.00
fryum	90.92	91.01	90.48	92.36	27.60	28.45	27.06	30.50	32.92	33.64	32.47	35.17

Appendix

Table 28: Per-class donor-forced counterfactual envelope, MVTec-AD

\to

VisA (same protocol and reading as Table 27 ).

Src cat.	Evolved normal rule	Evolved abnormal rule
candle	X clean intact candle	X visible anomaly on object
capsules	X a photo of a capsules	X plain damaged capsules
cashew	X a photo of healthy cashew	X faulty flawed cashew
chewinggum	X typical intact with chewinggum	X anomalous abnormal textural chewinggum
fryum	X a view of pristine good fryum	X item: defect in fryum
macaroni1	X a picture of macaroni1	X defective visible textural macaroni1

Appendix

Table 29: Evolved prompt rules per VisA source category, for the VisA

\to

MVTec-AD direction. Default

=

X normal/abnormal {cat} .

Src cat.	Evolved normal rule	Evolved abnormal rule
bottle	X normal bottle	X irregular visible surface bottle
cable	X typical cable	X an anomalous cable sample
capsule	X a standard capsule product	X a deformed structure capsule sample
carpet	X pristine surface carpet	X faulty blemished carpet
grid	X regular standard grid	X faulty flawed abnormal grid
hazelnut	X hazelnut	X abnormal faulty flawed hazelnut

Appendix

Table 30: Evolved prompt rules per MVTec-AD source category, for the MVTec-AD

\to

VisA direction. Default

=

X normal/abnormal {cat} .

Target	Nearest source	Sim	Route tag
bottle	candle	0.810	semantic (abn. fallback)
cable	macaroni1	0.758	semantic transfer
capsule	capsules	0.903	semantic transfer
carpet	fryum	0.799	semantic transfer
grid	fryum	0.804	semantic transfer
hazelnut	cashew	0.848	semantic transfer

Appendix

Table 31: VisA

\to

MVTec-AD transfer policy: MVTec-AD target

\to

VisA source. Sim

=

cosine similarity; Tag

=

policy tag.

Target	Nearest source	Sim	Route tag
candle	bottle	0.810	semantic transfer
capsules	capsule	0.903	semantic transfer
cashew	hazelnut	0.848	semantic transfer
chewinggum	toothbrush	0.716	semantic transfer
fryum	pill	0.819	semantic transfer
macaroni1	pill	0.826	semantic transfer

Appendix

Table 32: MVTec-AD

\to

VisA transfer policy: VisA target

\to

MVTec-AD source.

Figure A1: Analysis of selected rule embeddings. We report within-role cosine similarity and cross-role distance for the text embeddings of selected normal and abnormal rules.

Figure A2: Qualitative localization on MVTec-AD bottle . Top: input with ground-truth contours; bottom: CoEvoAD anomaly map.

Figure A3: Qualitative localization on MVTec-AD capsule .

Figure A4: Qualitative localization on MVTec-AD carpet .

Figure A5: Qualitative localization on MVTec-AD grid .

Figure A6: Qualitative localization on MVTec-AD hazelnut .

Figure A7: Qualitative localization on MVTec-AD leather .

Figure A8: Qualitative localization on MVTec-AD screw .

Figure A9: Qualitative localization on MVTec-AD tile .

Figure A10: Qualitative localization on MVTec-AD wood .

Figure A11: Qualitative localization on MVTec-AD zipper .

Figure A12: Qualitative localization on VisA candle .

Figure A13: Qualitative localization on VisA capsules .

Figure A14: Qualitative localization on VisA cashew .

Figure A15: Qualitative localization on VisA chewinggum .

Figure A16: Qualitative localization on VisA macaroni1 .

Figure A17: Qualitative localization on VisA pipe_fryum .

Figure A18: Qualitative localization on BTAD wood .

Figure A19: Qualitative localization on DAGM fabric2 .

Figure A20: Qualitative localization on KSDD2 metal .

Figure A21: Qualitative localization on RSDD metal15 .

Zero-Shot Anomaly Detection

Zero-shot anomaly detection finds unusual patterns in unseen domains without task-specific labeled examples.

See all 85 papers

EntroAD: Structural Entropy-Guided Prompt Adaptation for Zero-Shot Anomaly Detection
May 27, 2026Xinyu Zhao, Qingyun Sun, Jiayi Luo +1VLM Adaptation Medical Image Anomaly Detection
Beihang University Beijing, China
WMoE-CLIP: Wavelet-Enhanced Mixture-of-Experts Prompt Learning for Zero-Shot Anomaly Detection
Mar 6, 2026Peng Chen, Chao HuangPrompt Learning for VLMs Zero-Shot Anomaly Detection
School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen
VFAD: Variational Semantic Prompting Meets Frequency-Adaptive Representation Learning for Zero-Shot Anomaly Detection
Jul 31, 2026Peng Chen, Kaige Li, Wei Wang +5VLM Adaptation Frequency-Domain Feature Learning
School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen, China

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

EntroAD: Structural Entropy-Guided Prompt Adaptation for Zero-Shot Anomaly Detection

Xinyu Zhao, Qingyun Sun, Jiayi Luo, Jianxin Li

Beihang University Beijing, China

VLM Adaptation Medical Image Anomaly Detection Industrial Anomaly Detection Prompt Tuning Zero-Shot Anomaly Detection

Open paper AskPDF arXiv

Zero-Shot Anomaly Detection (ZSAD) aims to detect anomalies in unseen domains without target-domain adaptation. Recent CLIP-based methods have shown promising performance by leveraging prompt learning and visual-text alignment. However, most existing approaches rely on a single adaptation pathway, which may be insufficient for heterogeneous anomaly patterns across domains. In practice, anomalies exhibit vastly different characteristics, ranging from salient, localized structural disruptions to subtle, diffuse, and irregular variations. To address this challenge, we propose EntroAD, a structural entropy-guided zero-shot anomaly detection framework. Unlike previous methods, EntroAD introduces a dynamic routing mechanism to process different types of anomalies with specialized adaptation strategies. Specifically, we estimate patch-level structural entropy from self-attention-induced patch relations and use it as a proxy for relational uncertainty to guide anomaly-aware token routing. Based on this routing signal, we construct anomaly-aware routed tokens to better capture anomaly cues with different structural characteristics. We further introduce a confidence-aware dual-branch prompt adaptation module to stabilize visual-text alignment while preserving CLIP's transferable prior. Extensive experiments on 10 industrial and medical benchmarks show that EntroAD achieves state-of-the-art performance in challenging cross-dataset ZSAD settings.

Abstract

Zero-Shot Anomaly Detection

Explore similar work

EntroAD: Structural Entropy-Guided Prompt Adaptation for Zero-Shot Anomaly Detection

WMoE-CLIP: Wavelet-Enhanced Mixture-of-Experts Prompt Learning for Zero-Shot Anomaly Detection

VFAD: Variational Semantic Prompting Meets Frequency-Adaptive Representation Learning for Zero-Shot Anomaly Detection