GLAD, a computer-vision system for multi-view industrial anomaly detection, was reported as the top performer on nine of 10 measures in Real-IAD and all 10 measures in MANTA-Tiny. The paper reports scores at sample, image and pixel levels, covering both anomaly detection and pixel-level localization. These are benchmark results, not a direct measure of factory quality control.
The question behind the model is whether unrestricted cross-view information can hurt reconstruction-based anomaly detection. The paper defines cross-view information leakage as normal cues from intact views reaching the decoder and collapsing the normal-abnormal reconstruction gap. It asks whether explicitly restricted fusion can improve detection and localization.
Why the model limits view-to-view information
In this type of system, the reported comparison is between encoder outputs and decoder reconstructions. GLAD uses a frozen pretrained Transformer to encode the views, then sends them through a global, object-level fusion stage called OGA and a local, token-level stage called MMA. An MLP bottleneck and Transformer decoder complete the reconstruction path.
In OGA, the cross-attention output directly replaces the original patch tokens instead of being added as a residual. The original features are therefore discarded in that operation, which the paper describes as a way to avoid accumulating a normal template.
Two separate benchmark settings
Real-IAD covers 30 object classes and includes 99,721 normal and 51,329 abnormal images captured from five fixed viewpoints. It has eight annotated defect categories, with defect areas ranging from 0.1% to 6.75% of an image.
The paper describes MANTA as spanning 38 classes, more than 137.3K object instances and more than 8.6K abnormal samples. Its official MANTA-Tiny split contains 800 instances per class and 4,000 images per class across five views, for 152,000 images.
The reported rankings
On Real-IAD, GLAD was reported as best on nine of 10 individual metrics. At sample level, the three reported scores, AUROC, average precision and F1, were 92.4, 96.3 and 91.4. At image level, they were 89.7, 87.2 and 80.3. At pixel level, the four scores were 98.9, 44.7, 48.2 and 94.9, respectively, for AUROC, average precision, F1 and AUPRO.
Against Dinomaly, the reported Real-IAD pixel-level differences were 1.6 points in average precision, 0.9 in F1 and 0.9 in AUPRO. Against MVAD, the differences were 14.4 points in average precision and 11.4 in F1.
On MANTA-Tiny, GLAD was reported as best on all 10 metrics. Its sample-level scores were 92.7, 88.5 and 83.6; image-level scores were 92.0, 83.5 and 78.1; and pixel-level scores were 95.2, 47.6, 49.6 and 85.9.
Compared with Dinomaly, the reported MANTA-Tiny pixel-level gains were 2.8 points in average precision, 1.3 in F1 and 1.7 in AUPRO.
What the component comparisons found
The component ablation reports a +2.2-point change in sample-level AUROC for MMA alone on Real-IAD. The OGA comparison reported pixel-level changes of +2.0 points in average precision and +0.5 in AUPRO on Real-IAD, and +3.0 and +0.8 on MANTA-Tiny. The full model's P-AUPRO was 0.9 points higher than the no-module baseline on Real-IAD and 1.7 points higher on MANTA-Tiny.
A separate OGA restriction test compared four variants: token replacement, softmax-to-sigmoid gating, withholding the NoGlobal feedback path and a sharpened temperature of 0.7. The paper reports consistent gains on MANTA-Tiny in those comparisons, while sample-level measures stayed stable or improved alongside pixel-level measures.
Cost and missing views
The paper reports different resource figures for its fusion choices. Linear attention used 2.66 million parameters and 12.04 milliseconds per forward pass. MVAS used 3.86 million parameters and 47.51 milliseconds, with 9,276 megabytes of training memory compared with 1,312.9 megabytes for linear attention. Deformable attention used 2.37 million parameters and 3,091.1 megabytes.
The reported pixel-level vectors for linear and deformable attention were nearly comparable: 98.9/42.7/47.0/94.4 for linear attention and 98.9/43.3/46.7/94.4 for deformable attention.
In a targeted missing-view analysis, the authors masked views C2, C3 and C4, renormalized view-importance weights over the remaining views and did not retrain the model. After three of five views were removed, they reported largely stable pixel-level AUROC and sample-level AUROC above 90.
That result applies to the specified masking setup. It does not establish robustness to every possible pattern of missing cameras.
What remains uncertain
The findings remain limited to the two benchmark settings. For cross-method comparisons, the study reproduced single-view baselines under the multi-view setting, while MVAD was evaluated directly as a multi-view method. Detailed reproduction procedures are not provided in the supplied analysis.
The analysis reports no confidence intervals, hypothesis tests or run-to-run variability. It also reports no direct production deployment or quality-control outcome. The supplied evidence leaves open whether the rankings transfer to other datasets, camera arrangements and deployment conditions.
The item is identified as arXiv:2608.25168v1, dated 25 Aug 2026. The supplied front matter includes a CC BY 4.0 license statement, but no funding statement is reported in the supplied text or metadata.
Paper data and sources
Original title: See More, Detect Less? Taming Information Leakage in Multi-View Anomaly Detection
Authors: Shang-Fu Chen, Kuan-Chuan Peng, Jhih-Ciang Wu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text