Physical World & Real-World ImpactUniversity of Ljubljana (Fučka, Zavrtanik, Skočaj)
AnomalyVFM
Published AnomalyVFM — turns vision foundation models into zero-shot anomaly detectors through 3-stage synthetic data generation and LoRA. By the authors' measurement, average image-AUROC of 94.1% across 9 industrial benchmarks (previous best 90.8%), with 0 in-domain training images
Why it matters
The only approach that covers the early phase of a new product line, when there are not even normal images. However, there are no semiconductor- or PCB-specific results, so an in-house pilot is needed to check it
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/html/2601.20524v1Added in Week 35, 2026 · See that week