Domain generalization is an essential direction in deep learning, which mainly focuses on how to train a model on one or more source domains to perform well on unseen target domains, and the core key is in the model’s ability to generalize. In real-world applications, there are often specific differences between the train set (source domains) and the test set (target domains), and these differences may be due to different data collection times, different environmental conditions, different equipment, and other factors. In this paper, we present a framework that is grounded in the variational self-encoder, specifically, aligning the distributions of different domains by imposing the maximum mean difference and enhancing the training samples with Cutmix and then using variational self-encoder to learn the features in the domains in this way, we learn the standard features in each domain and thus extend the generalization ability of the model. Following experiments across diverse image tasks, our proposed approach demonstrates superior generalization capabilities compared to existing domain generalization techniques.