EdgeZSAD: Practical Zero-Shot Anomaly Detection on Edge Devices

A research team has introduced EdgeZSAD, a compact zero-shot anomaly detection system designed to run on edge devices, moving away from the large foundation models that dominate the field [1]. The system, detailed in a paper submitted on June 15, 2026, uses a TinyViT-21M-512 backbone, a model with 21 million parameters, a fraction of the roughly 300 million parameters found in the ViT-L foundation backbones used by recent methods [1][2]. This smaller footprint allows EdgeZSAD to be deployed directly on hardware such as the Jetson Orin Nano Super and RB5 Gen2 [1][2]. The model was trained using a single checkpoint in a source-trained, target-unseen protocol and evaluated across six industrial benchmarks [1][2]. Over three independent runs, it achieved an average image AUROC of 91.6 on the MVTec-AD dataset and 88.2 on the VisA dataset [1][2]. The system incorporates an asymmetric global-local readout, termed EdgeGLR, and a reproducible training recipe called Real-IAD-DR [1][2]. When the model was exported and rescored on the target edge devices, the image-AUROC drift remained below 0.2 points, suggesting the exported graph preserves host-side ranking behavior in the evaluated deployment setting [1][2]. The work was authored by Andrew Jaeyong Choi and colleagues [1]. The research addresses a gap in industrial inspection, where zero-shot anomaly detection must remain useful under the memory and operator constraints of embedded hardware [1][2].