Fault-Guided Cross-Layer Prediction for Unsupervised Anomaly Detection

Tao Wang

cs.CVNov 5, 2026

Fault-Guided Cross-Layer Prediction for Unsupervised Anomaly Detection

Authors: Tao Wang

Abstract

Feature reconstruction is widely used for unsupervised industrial anomaly detection in the single-class setting, where only normal images from one product category are available for training. We propose a single-class detector that asks what information one block can pass to another. We treat the depth of a frozen vision transformer as a sequence of discrete transitions. A directed predictor learns a block-to-block hidden-state transition pointwise over tokens. Adjacent-layer cosine similarity exposes four sharp fault lines in DINOv2 ViT-L, separating stable depth segments. The deepest blocks retain semantic context but lose local detail; a prediction from deep or shallow sources to sensitive targets therefore exposes the information that cannot be transported. The fault structure fixes the source and target roles, and the same rule transfers to ViT-B and RADIO-B. We use a bounded foreground prior calibrated solely on training images: it attenuates background evidence without also erasing defects outside the estimated foreground. Using one predictor architecture and a common training-only calibration protocol, image AUROC is 99.75 on MVTec-AD, 99.42 on VisA and 94.27 on Real-IAD. On Real-IAD the object-level AUROC is 96.26. At 256x256 pixel resolution, P-AUROC/P-AUPRO are 98.46/95.97 on MVTec-AD, 99.21/96.12 on VisA and 99.26/96.59 on Real-IAD.

Figures & tables

Figure 1: Measured evidence for fault-guided cross-layer prediction. ( a ) Adjacent-depth cosine similarity averaged over five categories. Four faults separate stable depth segments. L14 is marked as the midpoint of the stable segment and the participation-ratio peak. The two angular-residual manifolds show normal patches concentrating near the origin while defect patches spread outward for backward $21{\to}14$ and forward $1{\to}7$ . ( b ) One anomaly from each of four MVTec-AD categories. The first two rows use $21{\to}14$ and the last two use $1{\to}7$ , illustrating complementary backward and forward evidence; the adjacent column compares these with $14{\to}13$ and $7{\to}8$ under the same training run.

Figure 2: Pipeline overview. Training and inference share a frozen DINOv2 ViT-L/14 and eight directed predictors. Four cosine faults select six blocks; each MLP maps one source block to one target, and the loss averages the hardest 5% of tokens. The residual $r_{s\to t}=1-\cos(g_{s\to t}(f_{s}),f_{t})$ measures what the source cannot carry, and the predictor retains a trainable source shortcut $\alpha_{s\to t}f_{s}$ . At test time, a training-only soft prior weights the residuals. The image score uses $\max(z_{A},z_{B},wz_{G})$ after median/spatial baselines, smoothing, high-pass filtering and top-0.2% pooling. Pixel localization bypasses these image-level branches and is evaluated at $256\times 256$ resolution as the direct mean of the eight residual maps.

branch	source block $s$	target blocks $t$
forward	$1$	$14,\;7$
backward	$21,\;22,\;23$	$14,\;7$

Table 1: The eight directed channels. Both branches are trained jointly; the pairing rule is derived in Sec. 3.2 and the role of each direction is analysed in Sec. 3.3 .

backbone	detected faults	near target	mid target	deep source band
ViT-L	$0\|1$ , $7\|8$ , $21\|22$ , $22\|23$	$7$	$14$	$21,22,23$
ViT-B	$3\|4$ , $9\|10$	$3$	$7$	$10,11$
RADIO-B	$3\|4$ , $9\|10$	$3$	$7$	$10,11$

Table 2: Fault boundaries produced by the relative rule on the recorded cosine curves, together with the layer set they imply. The ViT-B and RADIO-B rows are the configurations used in the completed backbone-generalization experiments.

Figure 3: Source/target placement from layer-wise defect separation. Blue uses ground-truth masks; orange uses synthetic anomalies. Shading shows

\pm 1

standard deviation over five categories. L7 and L14 remain sensitive and serve as targets; separation falls over blocks 21–23, which serve as sources. L7 is kept as the near target because it is the last sensitive block before

7|8

, not because it is the global maximum.

quantity	symbol	value
high-pass strength	$\lambda$	$0.75\cdot\mathrm{clip}((\rho-0.80)/0.08,0,1)$
high-pass scale	$\sigma$	$2.0$
smoothing	–	$3\times 3$ Gaussian, $\sigma=0.8$
pooling ratio	–	top $0.2\%$ ( $k\geq 5$ )
B-branch field	$\sigma_{B}$	$4$
global reliability	$w$	$\exp(-\mathrm{rot}/0.30)$ ; $w=0$ for no mask or multi-object

Table 3: Fixed inference values and training-only calibration rules. Thresholds are shared;

\lambda

, normalisation and the mask prior use only normal training images.

method	S-AUROC	I-AUROC	P-AUPRO
PaDiM [ 10 ]	91.2	80.3	81.8
CFlow [ 14 ]	89.8	82.5	90.5
PatchCore [ 34 ]	93.7	89.4	91.5
SoftPatch [ 22 ]	92.8	89.4	90.8
SimpleNet [ 25 ]	94.9	88.5	84.6
DeSTSeg [ 41 ]	94.0	86.9	81.5

Table 4: Real-IAD reference benchmark. S, I and PRO are sample AUROC, image AUROC and P-AUPRO. The first eight rows use the official 256-side protocol; our patch-aligned

252^{2}

run is an approximate resolution match. This table is used for benchmark-level calibration, not as a strict protocol-identical claim.

		$392^{2}$				$448^{2}$
dataset	method	I	O	P	PRO	I	O	P	PRO
MVTec-AD	ours	99.70	–	97.30	95.94	99.75	–	98.46	95.97
MVTec-AD	Dinomaly2	99.84	–	98.54	95.83	99.85	–	98.75	96.42
VisA	ours	99.29	–	98.95	94.83	99.42	–	99.21	96.12
VisA	Dinomaly2	99.05	–	99.21	95.90	98.90	–	99.39	96.64
Real-IAD	ours	94.01	96.13	98.70	97.33	94.27	96.26	99.26	96.59

Table 5: Controlled Dinomaly2 comparison. I and O are image- and object-level AUROC; P and PRO are P-AUROC and P-AUPRO. Each pair uses the same resolution. Dinomaly2 uses its native center crop, whereas the refreshed ours run resizes the full image directly to

448^{2}

; its pixel maps and ground truth are compared on a common

256\times 256

grid. The

392^{2}

row is retained from the preceding matched protocol and is not part of the refreshed result.

configuration	mean AUROC	$\Delta$ (pp)
reference (8 predictors, B3)	0.9902	–
branch group
forward channels only	0.9663	$-2.39$
all backward channels (BWD6)	0.9874	$-0.28$
targets
drop target $L_{7}$ (ONLY14)	0.9866	$-0.36$

Table 6: Ablation on cable, screw, pcb1 and macaroni2 (mean image AUROC). Full channel sets and per-category results are in the supplement. The patch-resolution block uses six Real-IAD categories and is compared only within that subset.

backbone	MVTec-AD I	VisA I	Real-IAD I/O
ViT-L/14	99.75	99.42	94.27 / 96.26
ViT-B/14	99.46	99.04	94.01 / 96.02
RADIO-B	99.33	98.35	93.27 / 95.78

Table 7: Backbone generalization at

448^{2}

in image-level AUROC. The Real-IAD cell includes object-level AUROC (I/O). Pixel columns are omitted because their protocol was refreshed only for the final ViT-L run. The layer set changes only when the backbone changes.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

category	$\lambda$	$\rho$
bottle	0.000	0.7960
cable	0.000	0.7123
capsule	0.000	0.6818
carpet	0.000	0.5752
grid	0.000	0.6649
hazelnut	0.000	0.7785

Appendix

Table 8: MVTec-AD high-pass calibration.

category	$\lambda$	$\rho$
candle	0.000	0.6455
capsules	0.631	0.8673
cashew	0.000	0.7130
chewinggum	0.000	0.7926
fryum	0.604	0.8644
macaroni1	0.596	0.8635

Appendix

Table 9: VisA high-pass calibration.

category	$\lambda$	$\rho$
audiojack	0.651	0.8695
bottle_cap	0.000	0.7186
button_battery	0.745	0.8795
end_cap	0.530	0.8566
eraser	0.000	0.6983
fire_hood	0.295	0.8315

Appendix

Table 10: Real-IAD high-pass calibration.

dataset	I-AUROC	O-AUROC	P-AUROC	P-AUPRO
MVTec-AD	$99.75\pm 0.34$	–	$98.46\pm 1.22$	$95.97\pm 4.12$
VisA	$99.42\pm 0.53$	–	$99.21\pm 0.94$	$96.12\pm 1.98$
Real-IAD	$94.27\pm 3.51$	$96.26\pm 2.89$	$99.26\pm 1.12$	$96.59\pm 3.76$

Appendix

Table 11: Category-level mean

\pm

standard deviation for the production ViT-L/14 model. Standard deviations are computed over categories, not repeated runs.

configuration	forward channels	backward channels
Ours (8)	$f_{1\to 14},f_{1\to 7}$	$b_{21\to 14},b_{21\to 7},b_{22\to 14},b_{22\to 7},b_{23\to 14},b_{23\to 7}$
forward-only	$f_{1\to 14},f_{1\to 7}$	–
BWD4	–	$b_{21\to 14},b_{21\to 7},b_{22\to 14},b_{22\to 7}$
BWD6	–	$b_{21\to 14},b_{21\to 7},b_{22\to 14},b_{22\to 7},b_{23\to 14},b_{23\to 7}$
B1	$f_{1\to 14},f_{1\to 7}$	$b_{23\to 14},b_{23\to 7}$
B2	$f_{1\to 14},f_{1\to 7}$	$b_{22\to 14},b_{22\to 7},b_{23\to 14},b_{23\to 7}$

Appendix

Table 12: Exact source-to-target sets for the reported ablation configurations.

group	configuration	mean I-ROC	$\Delta$ (pp)
reference	NOINP	0.9902	–
control	F8NOINP	0.9901	$-0.02$
direction	forward-only	0.9663	$-2.39$
	BWD4	0.9868	$-0.34$
	BWD6	0.9874	$-0.28$
source	A0	0.9901	$-0.01$

Appendix

Table 13: Complete four-category ablation results. I-ROC is reported as a fraction;

\Delta

is relative to the matched no-input-injection reference (NOINP).

configuration	cable	screw	pcb1	macaroni2	mean
BWD6	0.9888	0.9904	0.9969	0.9733	0.9874
B1 (1 deep source)	0.9981	0.9873	0.9927	0.9711	0.9873
B2 (2 deep sources)	0.9972	0.9902	0.9970	0.9722	0.9892
B4 (4 deep sources)	0.9951	0.9918	0.9980	0.9758	0.9902

Appendix

Table 14: Per-category results for the new direction and deep-source-count ablations. B3 is the production reference (three deep sources) with mean

0.9902

. Adding a fourth source matches B3, so accuracy saturates at the production set.

category	$1{\to}14$	$1{\to}7$	best deep $\to$ 14	best deep $\to$ 7	mean
cable	0.9981	0.9875	0.9893 (23 $\to$ 14)	0.9920 (21 $\to$ 7)	0.9985
screw	0.9283	0.9524	0.9889 (22 $\to$ 14)	0.9902 (21 $\to$ 7)	0.9920
pcb1	0.9818	0.9889	0.9927 (23 $\to$ 14)	0.9862 (23 $\to$ 7)	0.9918
macaroni2	0.9286	0.9811	0.9907 (23 $\to$ 14)	0.9893 (23 $\to$ 7)	0.9915
audiojack	0.7852	0.8358	0.8485 (23 $\to$ 14)	0.8393 (21 $\to$ 7)	0.8486

Appendix

Table 15: Per-channel image AUROC on five diagnostic categories. “Best” is selected only for display within each category; the deployed model always uses all eight channels. The table illustrates channel specialization and is not used for per-category selection. The final mean column evaluates all eight channels together.

category	gate	prior off	prior on	$\Delta$ (pp)
cable	ON	0.9971	0.9985	$+0.14$
screw	OFF	0.9920	0.9920	$0.00$
pcb1	ON	0.9813	0.9918	$+1.05$
macaroni2	ON	0.9544	0.9915	$+3.71$

Appendix

Table 16: Image-level AUROC with and without the foreground prior. The prior is required for accepted masks and can be decisive when background evidence competes with a defect. The reliability gate keeps it disabled when the training mask is unreliable.

category	cross $21{\to}14$	recon $14$	grouped recon
cable	0.9648	0.9565	0.9515
screw	0.9553	0.9076	0.9129
pcb1	0.9658	0.9602	0.9630
macaroni2	0.9637	0.9806	0.9813

Appendix

Table 17: Mechanism-control pilot under matched head capacity and identical scoring. Cross-layer prediction wins on cable, screw and pcb1, while reconstruction is stronger on macaroni2. This is a four-category local pilot, not the full six-category study, and it motivates a complementary rather than universal claim.

category	patch-8	native-14	$\Delta$ (pp)
bottle_cap	97.84	93.54	$-4.30$
button_battery	93.83	88.46	$-5.37$
end_cap	91.89	88.17	$-3.72$
eraser	96.31	92.93	$-3.38$
fire_hood	93.48	91.71	$-1.77$
mounts	87.63	86.13	$-1.50$

Appendix

Table 18: Image-level AUROC for the patch-resolution control on six Real-IAD categories. patch-8 is the overlapping stride-8 tokenization; native-14 uses DINOv2’s native patch embedding. Removing the overlapping patchization lowers the matched mean by

3.34

pp.

dataset	category	I-AUROC	P-AUROC	P-AUPRO
MPDD	bracket_black	92.75	99.16	98.42
	bracket_brown	97.13	94.26	93.13
	bracket_white	99.78	99.59	94.86
	connector	100.00	97.82	93.17
	metal_plate	100.00	99.53	97.98
	tubes	99.95	99.10	96.82

Appendix

Table 19: Per-category results on the additional datasets. All values are percentages; pixel metrics use the same

256\times 256

evaluation grid as the refreshed main experiments.

dataset	categories	mask ON	gate OFF	no mask	$f=0.1$
MVTec-AD	15	7	3	5	2
VisA	12	11	1	0	9
Real-IAD	30	30	0	0	2

Appendix

Table 20: Mask-gate and floor assignments for the production ViT-L/14 runs. Gate-off categories are capsule, screw and transistor on MVTec-AD, and pcb3 on VisA. Real-IAD is gate-on for all 30 categories.

dataset	category	I (O)	P	PRO
MVTec-AD	bottle	100.00	99.20	97.70
	cable	99.55	98.69	95.79
	capsule	99.16	98.79	97.69
	carpet	100.00	99.43	98.35
	grid	100.00	99.44	96.78
	hazelnut	100.00	99.55	98.11

Appendix

Table 21: Per-category ViT-L/14 results for the refreshed

448^{2}

run. I, O, P and PRO denote image AUROC, object AUROC, P-AUROC and P-AUPRO.

Unsupervised Anomaly Detection

Unsupervised anomaly detection finds unusual patterns without labeled examples of faults, attacks, defects, or other rare events.

See all 233 papers

DuoAD: Leveraging [CLS] Dual Characteristics for Training-Free Few-Shot Anomaly Detection
Jul 27, 2026Jyun-Ze Tang, Po-Han Huang, Ming-Ching Chang +2Few-Shot Anomaly Detection Anomaly Localization
Inventec Corporation, Taipei, Taiwan · University at Albany, State University of New York, NY, USA
SwinAD: Multi-stage feature reconstruction for unsupervised industrial anomaly detection
Jul 16, 2026Huong Ninh, Chien Thai, Mai Xuan Trang +3Anomaly Localization Reconstruction-Based Anomaly Detection
Institute of Artificial Intelligence, University of Engineering and Technology, Vietnam National University · Applied AI Lab, Phenikaa School of Computing, Phenikaa University
UniVAD v2: Unified Visual Anomaly Detection via Support-Conditioned Boundary Construction
Jun 29, 2026Zhaopeng Gu, Bingke Zhu, Zhaowen Li +5Few-Shot Anomaly Detection Anomaly Localization
Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China · Yinwang Intelligent Technology Co. Ltd., China +4

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

DuoAD: Leveraging [CLS] Dual Characteristics for Training-Free Few-Shot Anomaly Detection

Jyun-Ze Tang, Po-Han Huang, Ming-Ching Chang, Chih-Fan Hsu+1

Inventec Corporation, Taipei, Taiwan · University at Albany, State University of New York, NY, USA

Few-Shot Anomaly Detection Anomaly Localization Anomaly Detection

Open paper AskPDF arXiv

Vision foundation models have enabled strong training-free anomaly detection (AD). However, most existing approaches rely primarily on independent local patch features, leaving the global contextual information encoded by Vision Transformers (ViTs) underexploited. In this work, we identify the dual characteristics of the ViT [CLS] token: its embedding provides anomaly-invariant global semantic representation, while its attention maps implicitly highlight spatially abnormal regions. Building on this observation, we propose a fully automated AD framework leveraging global context to remove manual tunings. Our framework introduces (1) an automatic augmentation selection strategy driven by [CLS]-level semantic consistency, and (2) an attention-guided feature reweighting mechanism that dynamically adjusts patch contributions according to [CLS] attention saliency. By integrating these components over multi-level features, our method achieves stable anomaly scoring and precise localization without training or parameter tuning. Under the one-shot setting, it achieves Image-AUC scores of 97.7%, 93.2%, and 84.5% on MVTec-AD, VisA, and Real-IAD. Using a single fixed configuration across categories, backbones, and datasets, the method establishes a new state-of-the-art for plug-and-play, training-free anomaly detection while maintaining strong robustness and practical scalability.

Abstract

Unsupervised Anomaly Detection

Explore similar work

DuoAD: Leveraging [CLS] Dual Characteristics for Training-Free Few-Shot Anomaly Detection

SwinAD: Multi-stage feature reconstruction for unsupervised industrial anomaly detection

UniVAD v2: Unified Visual Anomaly Detection via Support-Conditioned Boundary Construction