DifferAD-R1: A Difference-Guided IndustrialAnomaly Localization with Multimodal LargeLanguage Models

2026-06-15 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

DifferAD-R1 is an MLLM-augmented reinforcement learning framework designed for industrial anomaly localization, specifically targeting the detection of unseen defect categories in industrial products. It addresses the limitations of traditional closed-set methods, which struggle with cross-scenario generalization, and existing MLLM-based approaches that use misaligned QA-style paradigms or ineffective optimization for subtle defects. DifferAD-R1 introduces a Difference-Guided dual-image paradigm, reframing localization as a one-shot difference grounding problem to explore cross-scenario anomalies. It also features a Dual-Consistency Localization Reward to improve optimization stability for hard-to-detect anomalies and integrates a difficulty-aware strategy with adaptive reweighting and group-wise resampling. For evaluation, the AD-DualDiff dataset was constructed, comprising 13K paired images across 20 categories. Experimental results show DifferAD-R1 significantly outperforms existing baselines and achieves competitive performance against large-scale models like Qwen3-VL (235B parameters).

Key takeaway

For Computer Vision Engineers developing industrial anomaly detection systems, DifferAD-R1 offers a robust approach to overcome generalization issues and detect subtle defects. You should consider its Difference-Guided dual-image paradigm and Dual-Consistency Localization Reward to enhance your models' performance on unseen defect categories. Its difficulty-aware strategy can also improve learning efficiency on challenging instances, potentially reducing false negatives in critical industrial applications.

Key insights

DifferAD-R1 uses a difference-guided MLLM-augmented RL framework to localize industrial anomalies, improving generalization and detection of subtle defects.

Principles

• Reformulate localization as difference grounding.
• Enhance optimization for subtle defects.
• Prioritize learning on challenging instances.

Method

DifferAD-R1 employs a Difference-Guided dual-image paradigm, Dual-Consistency Localization Reward, and a difficulty-aware strategy with adaptive reweighting and group-wise resampling within an MLLM-augmented reinforcement learning framework.

In practice

• One-shot difference grounding for anomalies.
• Evaluate with 13K paired images.
• Publicly available code for implementation.

Topics

• Industrial Anomaly Localization
• Multimodal Large Language Models
• Reinforcement Learning
• Computer Vision
• Defect Detection
• AD-DualDiff Dataset

Code references

• Rong2026/work-1

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Rethinking Infrastructure Inspection as Image Difference Classification: A Traffic Sign Case Study — Image difference classification with instruction-based models reduces data dependency for infrastructure defect detection.
• DUALVISION: RGB-Infrared Multimodal Large Language Models for Robust Visual Reasoning — Integrating infrared and RGB data via localized cross-attention significantly boosts MLLM robustness in degraded visual conditions.
• OmniCD: A Foundational Framework for Remote Sensing Image Change Detection Guided by Multimodal Semantics — OmniCD unifies remote sensing change detection with multimodal semantic guidance, achieving robust, general-purpose performance across diverse scenarios.
• G2IA: Geometry-Guided Instance-Aware Retrieval and Refinement for Cross-Modal Place Recognition — G2IA improves cross-modal place recognition by combining geometry-guided retrieval with instance-aware refinement, overcoming modality gaps and perceptual aliasing.
• Visual Prompting Meets Feature Reconstruction-Based Anomaly Detection with Dual-Teacher Supervision — Robust anomaly detection in varied real-world conditions is achieved by combining visual prompting, adaptive teacher models, and synthetic data.
• Exploration Structure in LLM Agents for Multi-File Change Localization — Non-linear, domain-scoped parallel exploration significantly improves LLM agent performance for multi-file change localization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.