Towards Few-Shot Industrial Anomaly Generation under Object Misalignment: A Benchmark and a Method

Few-shot anomaly generation (FSAG) addresses the scarcity of anomalous samples by synthesizing realistic defects from a few anomaly references. However, existing FSAG methods are developed on object-aligned datasets, limiting their applicability to real industrial scenarios where objects often vary in position, scale, and quantity. Despite its practical importance, anomaly generation under object misalignment remains largely unexplored.To address this gap, we introduce the Misaligned Industrial Anomaly Dataset (MIAD), a benchmark for few-shot industrial anomaly generation and anomaly detection under object misalignment. MIAD contains $15,019$ normal images and $2,408$ anomalous image--mask pairs across $29$ object categories and $67$ anomaly types, covering diverse object layouts encountered in real-world inspection.We further propose AliGen, a plug-and-play framework that generates plausible anomaly masks for arbitrary normal images. AliGen introduces a texture-agnostic anomaly marker to encode anomaly locations and structural cues, and employs a Structure-Aware Diffusion Model (SADM) to produce spatially consistent anomaly masks conditioned on the input image. The generated masks can be seamlessly integrated into existing FSAG pipelines for realistic anomaly synthesis.Extensive experiments demonstrate that AliGen consistently improves both anomaly generation quality and downstream anomaly detection performance across multiple state-of-the-art FSAG methods.