This project aims to enhance the robustness, interpretability, and generalization of visual anomaly detection (VAD) systems in realistic, noisy environments. Building upon the adaptive deviation learning framework, we plan to implement an instance-wise reweighting strategy that dynamically adjusts the contribution of each sample during training, allowing the model to effectively handle data contamination. To improve explainability, we will integrate anomaly localization capabilities that highlight specific regions responsible for anomalous behavior, providing spatial insights alongside scalar anomaly scores. A key component of the project involves designing a pseudo-anomaly generation pipeline using diffusion models, where synthetic anomalies will be guided by structural prompts and domain priors to simulate realistic defects. These generated samples will serve as additional training signals to improve the model's ability to generalize to unseen anomalies. By combining adaptive learning, explainable outputs, and generative modeling, this project will deliver a comprehensive and scalable VAD framework. Extensive evaluation will be carried out on benchmark datasets such as MVTec and VisA, demonstrating improvements in detection accuracy, localization quality, and resilience to data contamination.