DeCoFlow: Structural Decomposition of Normalizing Flows for Continual Anomaly Detection

2026-06-25

AI Summary

This work addresses catastrophic forgetting in normalizing flow models when performing continual anomaly detection for sequentially arriving new product categories in industrial settings, where no historical data is available. To mitigate this issue, the authors propose a structured disentanglement approach that decomposes affine coupling layers into a frozen universal basis and task-specific low-rank adapters, thereby isolating parameters while strictly preserving model invertibility. The method integrates a task-alignment mechanism, auxiliary coupling layers, and a tail-aware loss to simultaneously achieve zero forgetting and high detection performance. Evaluated on MVTec-AD and VisA, the approach attains image-level AUROC scores of 98.40% and 93.00%, respectively, with a parameter-level forgetting rate of 0.00% and only 2.27M parameters per task.

Abstract

In industrial environments, new product categories arrive sequentially, requiring continual anomaly detection without access to past data. Normalizing Flows (NFs) provide exact density estimation but suffer from catastrophic forgetting as parameter updates across tasks distort the density manifold. While parameter isolation can prevent interference, it must preserve the strict invertibility and Jacobian validity of NFs. To satisfy these requirements, we exploit the inherent property that affine coupling layers maintain transformation validity regardless of subnet parameterization. Based on this, we propose DeCoFlow, which decomposes subnets into a frozen universal base and task-specific low-rank adapters to isolate updates. We further introduce Task-Specific Alignment, Auxiliary Coupling Layers, and Tail-Aware Loss to compensate for frozen-base rigidity. DeCoFlow achieves state-of-the-art image-level AUROCs of 98.40% on MVTec-AD and 93.00% on VisA, while maintaining parameter-level zero forgetting (0.00% FM under correct routing) with only 2.27M parameters per task.

Problem

• Continual Anomaly Detection
• Normalizing Flows
• Catastrophic Forgetting
• Industrial Inspection
• Density Estimation

Innovation

• Normalizing Flows
• Continual Anomaly Detection
• Low-Rank Adapters
• Catastrophic Forgetting
• Affine Coupling Layers