Quality-Aware Language-Conditioned Local Auto-Regressive Anomaly Synthesis and Detection

TL;DR

This paper introduces ARAS, a novel language-conditioned auto-regressive method for realistic local anomaly synthesis, integrated into QARAD, which significantly improves anomaly detection accuracy and speed over existing diffusion-based approaches.

Contribution

The paper presents ARAS, a new anomaly synthesis technique that enhances realism and control, and integrates it into QARAD for superior anomaly detection performance.

Findings

QARAD outperforms SOTA methods in benchmark datasets.

ARAS achieves 5x faster synthesis speed.

Enhanced defect realism and semantic control.

Abstract

Despite substantial progress in anomaly synthesis methods, existing diffusion-based and coarse inpainting pipelines commonly suffer from structural deficiencies such as micro-structural discontinuities, limited semantic controllability, and inefficient generation. To overcome these limitations, we introduce ARAS, a language-conditioned, auto-regressive anomaly synthesis approach that precisely injects local, text-specified defects into normal images via token-anchored latent editing. Leveraging a hard-gated auto-regressive operator and a training-free, context-preserving masked sampling kernel, ARAS significantly enhances defect realism, preserves fine-grained material textures, and provides continuous semantic control over synthesized anomalies. Integrated within our Quality-Aware Re-weighted Anomaly Detection (QARAD) framework, we further propose a dynamic weighting strategy that emphasizes high-quality synthetic samples by computing an image-text similarity score with a dual-encoder model. Extensive experiments across three benchmark datasets-MVTec AD, VisA, and BTAD, demonstrate that our QARAD outperforms SOTA methods in both image- and pixel-level anomaly detection tasks, achieving improved accuracy, robustness, and a 5 times synthesis speedup compared to diffusion-based alternatives. Our complete code and synthesized dataset will be publicly available.