TED:Text-Axis Evidence Decomposition for Prompted Anomaly Localization

JinYoung Kim; Geonho Kim; GiJeong Park; Geonu Lee; YoungJoon Yoo

cs.CVSep 30, 2026

TED:Text-Axis Evidence Decomposition for Prompted Anomaly Localization

Authors: JinYoung Kim, Geonho Kim, GiJeong Park, Geonu Lee, YoungJoon Yoo

Organizations: Department of Artificial Intelligence, Chung-Ang University, Seoul, Korea · SNUAILAB Co., Ltd., Seoul, Korea

Abstract

CLIP is a powerful vision-language model, but it was not designed for fine-grained defect localization; CLIP-based anomaly detectors therefore adapt it with prompts or lightweight modules to increase defect sensitivity. We show that stronger sensitivity does not necessarily make local evidence reliable: under domain shift, adapted CLIP-AD models often assign high anomaly scores to both true defects and visually complex normal regions. The issue is not simply missing defect information, but a local scoring rule that decodes defect and hard-normal evidence, having the same anomaly evidence. We propose TED (Text-Axis Evidence Decomposition), a post-hoc scoring method that asks whether each ambiguous response is better supported by source defect patches or by source normal patches mistaken as anomalous. TED compares these supports under the host's normal-versus-anomaly text response, leaves the backbone and prompts unchanged, and requires no target-domain training. It works as a train-free score for raw VLM backbones or as a source-calibrated residual correction for adapted CLIP-AD hosts. Across frozen VLM backbones, TED substantially improves pixel-level localization over raw prompt similarity; across adapted hosts, it improves most pixel-level settings over P-AUROC, P-PRO, and P-AP. Gains are largest under stronger hard-FP competition, with mean localization gain increasing from +5.0 in low-competition regimes to about +10.9 in mid/high-competition regimes. These results suggest that recoverable defect evidence can already exist in pretrained multimodal representations, but reliable localization requires decoding it against hard-normal competitors. Code will be released at TED GitHub repository.

Figures & tables

Panel of Figure 1 — Figure 1 : Hard-FP competition and text-axis response. (a) Across adapted CLIP-AD hosts, true defects remain much harder to separate from hard false positives than from generic outside patches. (b) Across target datasets, the learned text-axis response is consistently more informative than the orthogonal component, suggesting that the axis should be decomposed rather than discarded.

Figure 2 : Overview of TED . The host provides patch features( $V_{p}$ ) and normal $(t_{n})$ /anomaly $(t_{a})$ text embeddings that define the host text axis. TED projects query patches and source defect/hard-FP features onto this axis to compute a defect-versus-hard-normal support margin. T-TED uses this margin as a train-free local score, while C-TED converts it into a bounded residual anchored to the host score.

Panel of Figure 3 — Figure 3 : Text-response entanglement and source-supported re-ranking. Raw on-axis scoring strongly overlaps true defects with hard false positives. The orthogonal residual improves defect-vs-hard-FP separation in this case, but weakens separation from generic outside regions. The source-supported margin re-ranks the same responses while preserving strong outside separation, showing that TED decomposes rather than discards the text-guided response.

Figure 4 : Patch-level failure of raw prompt scoring. The left panel shows the raw prompt baseline and the right panel shows TED . The baseline ranks hard-normal patches close to or above true defects, whereas TED re-ranks them using source defect and hard-FP support without backbone or prompt adaptation. H-FP denotes hard false positive.

Figure 5 : Hard-FP decomposition in an adapted host. The baseline assigns overlapping high scores to true defect and hard-normal patches. TED re-scores the same responses with source defect and hard-FP evidence, reducing overlap and lifting the defect–hard-normal ranking gap.

Backbone	Variant	MVTec Bergmann et al. (2019)				VisA Zou et al. (2022)				MPDD Liu et al. (2024)				BTAD Mishra et al. (2021)				MVTec2 Heckler-Kram et al. (2026)
Backbone	Variant	I-AUC	P-AUC	PRO	AP	I-AUC	P-AUC	PRO	AP	I-AUC	P-AUC	PRO	AP	I-AUC	P-AUC	PRO	AP	I-AUC	P-AUC	PRO	AP
ViT-L/14-336	B	88.4	32.2	9.7	2.8	77.3	41.6	13.5	1.3	73.3	24.4	3.9	2.2	65.8	23.9	3.7	1.8	55.7	33.1	13.0	0.6
	T-TED	88.2	69.3	64.0	21.2	79.4	59.2	42.3	6.7	78.5	76.4	60.5	5.4	71.3	77.9	50.8	17.2	58.4	68.2	31.5	6.4
	C-TED	88.2	69.2	63.9	21.1	79.4	59.3	42.7	6.7	78.5	76.0	60.6	5.4	71.3	77.6	50.5	16.9	58.5	68.2	31.5	6.4
ViT-B/16+	B	89.5	17.3	1.7	2.4	73.2	21.7	3.4	0.4	60.1	28.0	4.1	1.4	67.1	38.3	7.5	2.2	54.7	35.5	5.7	0.5
	T-TED	90.0	82.9	67.8	17.1	75.5	77.4	51.9	6.1	59.4	70.8	48.4	10.2	69.0	63.3	34.8	6.1	54.5	64.5	29.7	2.7

Table 1: Cross-dataset transfer results for frozen VLM backbones with class-aware static prompts. B denotes standard prompt similarity, T-TED train-free TED, and C-TED source-calibrated TED. All variants use the same frozen backbone and prompts; only the local anomaly score changes. Best results within each backbone and target dataset are bolded.

Panel of Figure 6 — Figure 6 : Adapted-host and recipe-retuning diagnostics. Top : source-calibrated TED improves local pixel ranking across adapted CLIP-AD hosts; M, V, P, B, and M2 denote MVTec AD , VisA, MPDD, BTAD, and MVTec AD 2, respectively. Bottom : changing and retraining layer recipes shifts the host response, but the same traced defect and hard-FP patches remain competitive under the baseline. C-TED increases the patch-rank margin and yields cleaner aggregate separation between defect-supported and hard-normal-supported responses.

H-FP bin	#bins	Sev.	$\Delta$ PRO / $\Delta$ AP	$\Delta$ Loc
Low	3	0.07	+3.7 / +6.3	+5.0
Mid	3	0.10	+8.8 / +13.8	+11.3
High	3	0.19	+11.3 / +9.8	+10.6

Table 2 : Failure-conditioned gains. Bins are grouped by baseline hard-FP severity before inspecting C-TED gains.

Panel of Figure 7 — Figure 7 : Recipe-consistent retuning and hard-FP support. Top : C-TED remains beneficial after changing the feature-layer or backbone recipe, retraining the host when applicable, and rebuilding source banks for the resulting representation. Bottom : defect-only support leaves defect and hard-normal patches entangled, whereas adding hard-FP support explicitly models the competing hard-normal evidence and produces a cleaner local ranking margin.

Host	Method	P-AUROC	P-PRO	P-AP
AA-CLIP	Baseline	92.48	59.78	36.55
	T-TED	86.58	53.50	27.88
	C-TED w/o hard-FP	93.58	61.76	38.40
	C-TED	94.23	65.42	39.68
FAPrompt	Baseline	90.52	60.32	21.54
	T-TED	90.00	60.12	20.33

Table 3 : Ablation study of TED on BTAD. (a) validates the calibrated decoding components on adapted hosts, and (b) analyzes the effect of the residual rank on AA-CLIP.

Control	$n$	$\Delta$ PRO / $\Delta$ AP	$\Delta$ Loc
Scalar replace	16	-19.8 / -14.1	-17.0
Scalar residual	16	+2.1 / -5.3	-1.6
C-TED	16	+4.4 / +3.7	+4.1

Table 4 : Source-control ablations. Controls corrupt source supervision or replace TED with scalar source readouts using the same evidence. Values are gains over the host baseline.

Bank	Ret.	$\Delta$ PRO	$\Delta$ AP	$\Delta$ Loc
64	256	+7.17	+13.33	+10.25
128	512	+7.12	+13.32	+10.22
512	2048	+7.18	+13.33	+10.25

Table 5 : Fixed-bank sensitivity.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Item	Policy
Defect bank	Source anomalous patches overlapping source masks
Hard-FP bank	High-scoring source normal patches under the host baseline
Target usage	Target masks and scores are used only for final evaluation
Bank budget	Fixed per host; not selected per target dataset
Hard-FP mining	Fixed source-normal top fraction
Support bandwidth	Fixed before target evaluation

Appendix

Table 6 : Fixed evaluation policy. All entries are fixed before target evaluation and are not selected using target-domain scores or masks.

Host / transfer	Control	$\Delta$ P-AUC	$\Delta$ PRO	$\Delta$ AP
AA-CLIP MVTec $\rightarrow$ BTAD	C-TED	+1.03	+3.88	+1.71
AA-CLIP MVTec $\rightarrow$ BTAD	Label swap	-6.40	-8.28	-10.45
AA-CLIP VisA $\rightarrow$ BTAD	C-TED	-0.98	+0.69	+4.09
AA-CLIP VisA $\rightarrow$ BTAD	Label swap	-12.19	-15.95	-17.23
AdaCLIP VisA $\rightarrow$ BTAD	C-TED	+0.31	+2.84	+0.41
AdaCLIP VisA $\rightarrow$ BTAD	Label swap	-0.29	+0.45	-0.28

Appendix

Table 7 : Training-control sanity checks for C-TED. All variants use the same source-only calibration budget and residual capacity. Panel (a) tests the effect of swapping defect and hard-normal supervision across host/transfer settings. Panel (b) compares additional corrupted-control variants on AA-CLIP MVTec

\rightarrow

BTAD. Values report gains over the corresponding host baseline.

Figure 8 : Source-supervision controls for C-TED. All variants use the same source split, source-bank size, residual capacity, and calibration budget. Correct C-TED preserves the defect-versus-hard-FP assignment, while label shuffle and label swap corrupt only the supervision semantics.

Figure 9 : Raw VLM source-support ablation. Raw prompt scoring is compared with source-supported readouts in frozen VLM backbones. Text-axis support projects query and source-bank features onto the normal-versus-anomaly text axis, while visual support uses the source banks.

Host	Layers	M	I-AUC	P-AUC	P-PRO	P-AP
AdaCLIP	full	B	69.01	96.09	30.48	29.76
		T	69.46	96.08	30.67	29.70
		C	69.25	96.34	44.00	29.93
	last	B	61.59	84.31	40.79	24.29
		T	70.00	84.31	43.06	24.31
		C	70.42	89.62	54.00	25.95

Appendix

Table 8 : Layer-recipe stress and TED recovery on MPDD dataset. B, T, and C denote the baseline host, train-free TED, and source-calibrated TED.

Host	Budget	$n$	Retained	ms/img	Mem.	$\Delta$ Loc
AA-CLIP	64	6	6634.0	206.8	3.8GB	2.7
AA-CLIP	128	6	12180.0	182.9	3.8GB	2.9
AA-CLIP	512	6	35238.0	149.3	3.9GB	2.9
FAPrompt	64	6	256.0	1070.1	4.1GB	9.1
FAPrompt	128	6	512.0	1041.9	4.1GB	8.6
FAPrompt	256	6	1024.0	1055.2	4.1GB	9.1

Appendix

Table 9 : Fixed source-bank budget sensitivity. Values are averaged over completed transfer settings for each host and bank budget.

\Delta

Loc is the mean of

\Delta

PRO and

\Delta

AP over the host baseline.

Host	Transfer	$\Delta$ P-AUC	$\Delta$ PRO	$\Delta$ AP
AA-CLIP	MVTec $\to$ BTAD	-5.75	-7.53	-11.35
	MVTec $\to$ MPDD	+0.58	+4.92	+3.27
	MVTec $\to$ VisA	-0.43	-1.58	-3.28
	VisA $\to$ BTAD	-13.48	-16.22	-27.88
	VisA $\to$ MPDD	-2.53	-6.18	-5.98
	VisA $\to$ MVTec	-5.45	-6.30	-12.33

Appendix

Table 10 : Transfer-wise full-feature memory replacement control. The control uses the same source visual banks but removes the text-conditioned readout. Values are gains over the corresponding host baseline.

Readout	Mode	$\Delta$ P-AUC	$\Delta$ P-PRO	$\Delta$ P-AP
Full-D 1-NN	Replacement	$-45.65$	$-39.39$	$-20.68$
	Residual	$-10.99$	$-0.22$	$-7.66$
Full-D prototype	Replacement	$-11.85$	$-3.13$	$-12.63$
	Residual	$-3.27$	$+11.04$	$-5.64$
Full-D LogMeanExp	Replacement	$-24.70$	$-18.56$	$-16.67$
	Residual	$-8.68$	$+3.69$	$-7.86$

Appendix

Table 11 : Extended same-bank retrieval controls across AA-CLIP, AdaCLIP, and FAPrompt, with six transfers per host (18 settings in total). Values are mean changes in percentage points relative to the corresponding host. Source banks, hosts, targets, and aggregation are fixed; only the evidence readout changes. Replacement substitutes the host local score, whereas residual mode adds a correction to it.

Host	Source budget	$n$	$\Delta$ PRO	$\Delta$ AP	$\Delta$ Loc
AA-CLIP	1 img/cls	4	+3.3	+1.9	+2.6
AA-CLIP	2 img/cls	4	+3.5	+1.7	+2.6
AA-CLIP	4 img/cls	4	+3.2	+3.2	+3.2
AA-CLIP	all img/cls	4	+3.1	+2.9	+3.0
FAPrompt	8 bank/cls	3	+7.2	+13.4	+10.3
FAPrompt	16 bank/cls	3	+7.3	+13.4	+10.4

Appendix

Table 12 : Weak-source robustness. AA-CLIP varies the number of source images per source class used to construct banks. FAPrompt varies the retained source-bank budget per class. Values are gains over the corresponding host baseline;

\Delta

Loc is the mean of

\Delta

PRO and

\Delta

AP.

Host	$n$	Replace	Residual	C-TED
AA-CLIP	4	-7.4	-4.7	+3.0
AdaCLIP	4	-12.3	-4.1	+4.1
FAPrompt	4	-42.9	+3.8	+10.0

Appendix

Table 13 : Source-supervised readout control. The control uses the same source defect and hard-FP banks as C-TED, but replaces the evidence comparison with a directly learned MLP rank readout. Replace uses the learned readout as the local anomaly map, while Residual adds it to the host map. Rows average completed transfer settings for each host. Values are

\Delta

Loc, the mean of

\Delta

PRO and

\Delta

AP over the host baseline.

$\alpha$	0	0.25	0.5	1	2
$\Delta$ Loc	+0.0	+0.3	+0.3	-0.0	-5.2

Appendix

Table 14 : Residual-strength boundary. AA-CLIP diagnostic with fixed calibrator and protocol.

Host	$n$	Host replace	Host residual	C-TED
AA-CLIP	4	$+0.1/-4.5$	$-2.3/-7.7$	$\mathbf{+3.1/+2.9}$
FAPrompt	4	$+3.0/+3.3$	$+4.3/+5.0$	$\mathbf{+6.4/+13.5}$
AdaCLIP	4	$+5.3/-9.7$	$+0.1/-5.9$	$\mathbf{+7.4/-0.3}$

Appendix

Table 15 : Host-score-only calibration control. Controls train a source-only rank readout using only the scalar host local score on source defect and hard-FP patches, without source defect/hard-FP support features. Each entry reports

\Delta

PRO /

\Delta

AP over the corresponding host baseline.

Figure 10 : Raw VLM failure and TED recovery. Raw prompt scoring confuses defect evidence with hard-normal saliency, while TED recovers a cleaner local ranking using source-supported evidence.

Figure 11 : Qualitative comparisons of FAPrompt and C-TED across four target datasets, with ground-truth defects outlined in red and a shared color scale for each Host/C-TED pair.

Scope	$n$	$\Delta$ AUC [95% CI]	$\Delta$ Loc [95% CI]
FAPrompt	36	+8.1 [+6.3, +10.0]	+8.4 [+7.1, +9.7]
AA-CLIP	10	+2.2 [-1.9, +5.9]	+11.3 [-2.0, +25.1]
AdaCLIP	12	-2.5 [-4.3, -1.1]	+4.6 [+3.5, +5.7]
AdaptCLIP	12	+1.2 [+1.1, +1.3]	+0.3 [+0.1, +0.5]
All adapted hosts	70	+4.3 [+2.8, +5.8]	+6.8 [+4.7, +9.1]

Appendix

Table 16 : Failure-conditioned separability robustness. Intervals are setting-level percentile bootstrap 95% CIs over completed adapted-host settings.

\Delta

AUC measures the change in AUC for ranking anomalous regions above hard normal regions.

\Delta

Loc is the mean of

\Delta

PRO and

\Delta

AP.

Figure 12 : Failure-conditioned class-level cases. Representative target classes where C-TED improves localization while also increasing defect-versus-hard-FP AUC and the defect-minus-hard-FP score gap.

Host	Backbone	Target	Method	I-AUROC	P-AUROC	P-PRO	P-AP
AA-CLIP	ViT-B/16+	MVTec	Base	50.81 $\pm$ 0.00	39.27 $\pm$ 0.00	11.86 $\pm$ 0.00	2.97 $\pm$ 0.00
		MVTec	OURS	60.67 $\pm$ 0.05	69.22 $\pm$ 0.35	38.88 $\pm$ 0.82	10.39 $\pm$ 0.17
		VisA	Base	46.61 $\pm$ 0.00	54.49 $\pm$ 0.00	22.32 $\pm$ 0.00	1.31 $\pm$ 0.00
		VisA	OURS	54.41 $\pm$ 0.14	63.66 $\pm$ 0.22	28.74 $\pm$ 0.29	2.18 $\pm$ 0.04
		MPDD	Base	44.86 $\pm$ 0.00	88.93 $\pm$ 0.00	65.37 $\pm$ 0.00	12.10 $\pm$ 0.00
		MPDD	OURS	48.04 $\pm$ 0.21	88.98 $\pm$ 0.28	65.44 $\pm$ 0.67	12.07 $\pm$ 0.05

Appendix

Table 17 : AA-CLIP with host-anchored residual subspace correction. Results are averaged over three seeds; each entry is mean

\pm

std .

Host	Backbone	Target	Method	I-AUROC	P-AUROC	P-PRO	P-AP
FAPrompt	ViT-B/16+	MVTec	Base	80.75 $\pm$ 0.00	65.61 $\pm$ 0.00	35.51 $\pm$ 0.00	8.78 $\pm$ 0.00
		MVTec	OURS	80.75 $\pm$ 0.00	77.95 $\pm$ 0.00	62.34 $\pm$ 0.01	24.52 $\pm$ 0.00
		VisA	Base	69.16 $\pm$ 0.00	67.98 $\pm$ 0.00	40.10 $\pm$ 0.00	3.61 $\pm$ 0.00
		VisA	OURS	69.16 $\pm$ 0.00	71.56 $\pm$ 0.01	54.98 $\pm$ 0.08	8.79 $\pm$ 0.00
		MPDD	Base	67.35 $\pm$ 0.00	81.33 $\pm$ 0.00	59.09 $\pm$ 0.00	11.07 $\pm$ 0.00
		MPDD	OURS	67.35 $\pm$ 0.00	82.45 $\pm$ 0.01	61.23 $\pm$ 0.17	12.98 $\pm$ 0.01

Appendix

Table 18 : Results with FAPrompt as host.

Host	Backbone	Target	Method	I-AUROC	P-AUROC	P-PRO	P-AP
AdaptCLIP	ViT-B/16+	MVTec	Base	66.05 $\pm$ 4.16	68.91 $\pm$ 10.44	43.40 $\pm$ 9.63	15.22 $\pm$ 3.02
		MVTec	OURS	66.06 $\pm$ 4.14	68.63 $\pm$ 10.58	42.75 $\pm$ 9.76	15.41 $\pm$ 3.01
		VisA	Base	54.46 $\pm$ 3.63	50.17 $\pm$ 14.41	22.72 $\pm$ 16.49	2.40 $\pm$ 2.06
		VisA	OURS	54.21 $\pm$ 3.27	50.01 $\pm$ 14.65	22.33 $\pm$ 16.42	2.36 $\pm$ 1.99
		MPDD	Base	57.00 $\pm$ 3.18	69.97 $\pm$ 7.55	35.12 $\pm$ 5.41	6.10 $\pm$ 2.76
		MPDD	OURS	57.01 $\pm$ 3.18	70.20 $\pm$ 7.70	35.51 $\pm$ 5.62	6.10 $\pm$ 2.80

Appendix

Table 19 : Results with AdaptCLIP as host.

Host	Backbone	Target	Method	I-AUROC	P-AUROC	P-PRO	P-AP
AdaCLIP	ViT-B/16+	MVTec	Base	48.53 $\pm$ 7.16	50.32 $\pm$ 6.37	19.74 $\pm$ 5.72	4.27 $\pm$ 0.97
		MVTec	OURS	49.67 $\pm$ 8.24	50.26 $\pm$ 6.45	20.81 $\pm$ 5.70	4.46 $\pm$ 1.17
		VisA	Base	49.19 $\pm$ 1.25	57.15 $\pm$ 7.57	25.60 $\pm$ 6.46	1.16 $\pm$ 0.35
		VisA	OURS	50.83 $\pm$ 1.63	58.70 $\pm$ 6.52	28.22 $\pm$ 5.92	1.32 $\pm$ 0.39
		MPDD	Base	43.82 $\pm$ 2.60	42.46 $\pm$ 10.95	14.05 $\pm$ 6.18	2.82 $\pm$ 0.87
		MPDD	OURS	42.64 $\pm$ 0.32	42.57 $\pm$ 9.39	15.00 $\pm$ 5.11	2.61 $\pm$ 0.71

Appendix

Table 20 : Results with AdaCLIP as host.

Host	Backbone	Target	Method	I-AUROC	P-AUROC	P-PRO	P-AP
BayesPFL	ViT-B/16+	MVTec	Base	49.68 $\pm$ 1.01	49.73 $\pm$ 3.18	17.14 $\pm$ 3.94	4.29 $\pm$ 0.80
		MVTec	OURS	52.49 $\pm$ 2.56	49.67 $\pm$ 3.19	17.00 $\pm$ 3.86	4.26 $\pm$ 0.78
		VisA	Base	51.41 $\pm$ 1.85	54.71 $\pm$ 3.85	18.93 $\pm$ 5.43	0.90 $\pm$ 0.26
		VisA	OURS	50.20 $\pm$ 2.31	54.60 $\pm$ 3.67	18.87 $\pm$ 5.32	0.88 $\pm$ 0.24
		MPDD	Base	50.35 $\pm$ 4.02	58.47 $\pm$ 9.08	22.52 $\pm$ 9.48	2.39 $\pm$ 0.29
		MPDD	OURS	51.26 $\pm$ 4.88	58.27 $\pm$ 8.73	22.43 $\pm$ 9.22	2.38 $\pm$ 0.27

Appendix

Table 21 : Results with BayesPFL as host using calibrated TED.

Transfer	Host	Recipe	B	T	C
VisA $\rightarrow$ MVTec	AdaCLIP	Full(B)	89.93 / 89.89 / 44.03 / 41.60	91.43 / 89.88 / 43.60 / 41.46	91.64 / 89.86 / 47.97 / 41.85
		EL-r	87.83 / 80.21 / 18.18 / 22.40	88.74 / 80.21 / 17.94 / 22.39	88.84 / 80.96 / 26.25 / 23.81
		M2-r	89.26 / 87.62 / 23.50 / 36.18	87.74 / 87.61 / 23.63 / 36.12	87.74 / 87.79 / 25.42 / 36.40
		L2-r	90.58 / 89.47 / 15.19 / 36.35	89.07 / 89.47 / 15.14 / 36.31	88.69 / 90.84 / 22.78 / 39.28
		L-r	90.40 / 88.96 / 34.83 / 35.78	91.99 / 88.96 / 34.66 / 35.77	91.79 / 89.15 / 46.27 / 36.77
	FAPrompt	Full(B)	91.45 / 89.95 / 73.57 / 22.35	91.45 / 40.41 / 25.32 / 8.16	91.45 / 90.19 / 76.83 / 30.94

Appendix

Table 22 : Bidirectional layer-recipe retraining diagnostic. Each cell reports I-AUC / P-AUC / PRO / AP. B, T, and C denote the retrained host baseline, train-free TED, and source-calibrated TED, respectively. Full(B) is the official full-layer baseline recipe; EL-r, M2-r, L2-r, and L-r denote checkpoints retrained with early+last, middle-2, last-2, and last-layer recipes.

Explore similar work

Dec 10, 2025cs.CV

Defect-aware Hybrid Prompt Optimization via Progressive Tuning for Zero-Shot Multi-type Anomaly Detection and Segmentation

Recent vision-language models (VLMs) like CLIP have shown impressive anomaly detection performance under significant distribution shift by utilizing high-level semantic information through text prompts. However, these models often overlook fine-grained defect cues, e.g., hole, cut, or scratch, that are essential for understanding the anomaly's nature. Moreover, the modality gap between images and text can lead to subtle visual evidence being poorly captured in textual descriptions. To address the gap, we enhance the representation of "abnormal" with structured semantics, bridging coarse anomaly signals and fine-grained defect categories. We propose a hybrid prompting mechanism that combines human-readable descriptions of defect types with learnable token embeddings. Building on these ideas, we introduce DAPO, a Defect-aware Prompt Optimization framework for zero-shot multi-type and binary anomaly detection and segmentation under distribution shift. DAPO aligns anomaly-relevant visual features with their corresponding textual semantics by learning hybrid defect-aware prompts that combine fixed textual anchors with trainable token embeddings. We conducted experiments on public benchmarks (MPDD, VisA, MVTec-AD, MAD, and Real-IAD) and an internal dataset. The results suggest that compared to the baseline models, DAPO achieves a 3.6% average improvement in AUROC and average precision metrics at the image level under distribution shift, and a 5.2% average improvement in AUROC and F1 when localizing novel anomaly types under zero-shot settings.

Nadeem Nazer, Hongkuan Zhou, Lavdim Halilaj +2

Corporate Research, Robert Bosch GmbH, Renningen, Germany · Otto-von-Guericke-University Magdeburg, Magdeburg, Germany · Institute for Artificial Intelligence, University of Stuttgart, Stuttgart, Germany +2

Multimodal Anomaly Detection Few-Shot Segmentation Modality Gap Tuning

Ask PDF

Sep 7, 2026cs.CV

Proximity-CLIP: Text-Guided Semantic Proximity Learning for Zero-Shot Anomaly Detection

Vision-language models offer a promising approach for zero-shot anomaly detection (ZSAD). However, due to object-centric bias, normal and anomalous text prototypes exhibit a high semantic overlap. While enforcing strict orthogonality between them improves discriminability, mapping highly contiguous visual inputs onto drastically orthogonal prototypes introduces a geometric dilemma, disrupting the pre-trained structural continuity. To address this problem, we propose Proximity-CLIP, a framework that visually calibrates the semantic margin to guide visual adaptation. First, we introduce a visually-calibrated semantic proximity learning mechanism that uses a bounded dynamic regularization to learn an appropriate semantic margin, ensuring discriminative separation while preserving structural alignment. Second, we design an Anomaly Query Module (AQM) driven by these text priors. Using the calibrated anomalous prototype as a semantic query, the AQM actively retrieves localized defect cues from contextual visual patches, mitigating the dilution of subtle anomalies during global pooling. Extensive experiments demonstrate that Proximity-CLIP outperforms current state-of-the-art methods across multiple ZSAD benchmarks with minimal architectural modifications.

Manwen Yang, Leqian Ding, Yu Guo +1

School of Software Engineering, Xi’an Jiaotong University · State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University

Abnormality Zero-Shot Proximity Operations Geometric Bias

Ask PDF

Sep 15, 2026cs.CV

PSMP-CLIP: Patch-Prompt SAM and Multi-Semantic Prompting for CLIP-Based Zero-Shot Anomaly Detection

Zero-shot anomaly detection aims to localize anomalies without target-domain samples. Existing CLIP-based methods suffer from coarse anomaly maps and limited semantic prompts. We propose PSMP-CLIP, integrating patch-prompt SAM2 segmentation (PPSS) and multi-semantic guided prompt regularization (MSGPR). PPSS samples prompts directly from intermediate patch features, avoiding threshold drift and guiding SAM2 to produce precise masks. MSGPR uses multiple learnable prompts constrained by semantic anchors to preserve generalization. Experiments on 14 datasets show highly competitive performance, achieving the best pixel-level AUROC on MVTec AD, BTAD, DTD-Synthetic, CVC-ClinicDB, TN3K, Endo, and Kvasir.

Xuezhi Xiang, Guanghao Wu, Heqi Xiang +4

Information and Communication Engineering, Harbin Engineering University, Harbin, China · Key Laboratory of Advanced Marine Communication and Information Technology, Harbin, China · Department of Computer Science, University of Toronto, Toronto, ON M5S 2E4, Canada +1

Multimodal Anomaly Detection Few-Shot Segmentation Multiple Prompt Learning Frameworks Zero-Shot Segment Anything Model Regularization

Ask PDF