DifferAD-R1: A Difference-Guided Industrial Anomaly Localization With Multimodal Large Language Models

Abstract

Industrial anomaly localization aims to accurately identify and localize abnormal regions in industrial products, addressing the critical challenge of detecting unseen defect categories in real-world scenarios. Traditional closed-set methods often suffer from poor cross-scenario generalization, while existing Multimodal Large Language Model (MLLM)-based approaches face two core limitations: they either adopt QA-style paradigms misaligned with the practical demands of localization, or rely on standard optimization techniques such as Group Relative Policy Optimization (GRPO), which fails to deliver effective learning signals for subtle defects. To tackle these issues, this paper proposes DifferAD-R1, an MLLM-augmented reinforcement learning framework tailored for industrial anomaly localization. We design a Difference-Guided dual-image paradigm, which reformulates the localization task as a one-shot difference grounding problem to effectively explore cross-scenario anomalies. A Dual-Consistency Localization Reward (DCLR) is developed for hard-to-detect anomalies, enhancing optimization stability and robustness. Additionally, we integrate a difficulty-aware strategy with adaptive reweighting and group-wise resampling to prioritize learning on challenging instances. To facilitate evaluations in real-world industrial settings, we construct the AD-DualDiff dataset, comprising 13K paired images across 20 categories. Experimental results demonstrate that DifferAD-R1 significantly outperforms existing baselines and achieves competitive performance compared to large-scale models like Qwen3-VL (235B parameters). Our code is publicly available at: https://github.com/Rong2026/work-1