The paper introduces Real-IAD, a large-scale, real-world, and multi-view Industrial Anomaly Detection dataset. Real-IAD contains 150K high-resolution images of 30 different objects, significantly larger than existing datasets like MVTec AD and VisA. The dataset addresses the limitations of these datasets by providing a more diverse range of defect areas and ratios, making it more challenging for anomaly detection methods. To make the dataset closer to real-world applications, the authors adopt a multi-view shooting method and propose sample-level evaluation metrics. Additionally, they introduce a new setting called Fully Unsupervised Industrial Anomaly Detection (FUIAD), which is more practical for industrial settings where the yield rate is typically over 60%. The paper reports the performance of popular IAD methods on Real-IAD, providing a benchmark to promote the development of the field. The results show that while existing methods perform well on existing datasets, they still have room for improvement on Real-IAD, highlighting the dataset's challenge level. The main contributions of the paper include the proposal of Real-IAD, the construction of FUIAD, and the benchmarking of IAD methods on Real-IAD.The paper introduces Real-IAD, a large-scale, real-world, and multi-view Industrial Anomaly Detection dataset. Real-IAD contains 150K high-resolution images of 30 different objects, significantly larger than existing datasets like MVTec AD and VisA. The dataset addresses the limitations of these datasets by providing a more diverse range of defect areas and ratios, making it more challenging for anomaly detection methods. To make the dataset closer to real-world applications, the authors adopt a multi-view shooting method and propose sample-level evaluation metrics. Additionally, they introduce a new setting called Fully Unsupervised Industrial Anomaly Detection (FUIAD), which is more practical for industrial settings where the yield rate is typically over 60%. The paper reports the performance of popular IAD methods on Real-IAD, providing a benchmark to promote the development of the field. The results show that while existing methods perform well on existing datasets, they still have room for improvement on Real-IAD, highlighting the dataset's challenge level. The main contributions of the paper include the proposal of Real-IAD, the construction of FUIAD, and the benchmarking of IAD methods on Real-IAD.