Flownomaly: Single-Step Conditional-Mean Recovery for Few-Shot Anomaly Detection

TL;DR

Flownomaly introduces a single-step conditional-mean recovery method for few-shot anomaly detection by learning a velocity field in frozen DINOv2 feature space to map noisy query tokens to normal counterparts. It avoids trajectory integration and anomalous validation by using a normal-only Jacobian gain-to-noise criterion to set recovery time, enabling high-precision anomaly localization and detection.

Why It Matters

The method achieves state-of-the-art performance in extreme few-shot settings without requiring anomalous data for tuning, making it highly practical for industrial applications where defect samples are scarce or unavailable. Its single-step inference also enables fast deployment in real-time anomaly detection systems.

Key Findings

• Single-step recovery via conditional flow matching eliminates the need for trajectory integration during inference.
• A normal-only Jacobian gain-to-noise criterion enables optimal recovery time selection without anomalous validation data.
• Flownomaly achieves 97.8% image-level AUROC and 98.0% pixel-level AUROC on 1-shot MVTec-AD, outperforming prior methods.