We propose a novel AD pretraining framework specially designed for learning robust and discriminative pretrained representations for industrial anomaly detection.