The improvement of DNN model generalization relies on the availability of datasets characterized by a multitude of samples, balanced class distribution, and rich feature diversity. However, obtaining such datasets is a formidable challenge, and the problem of inadequate diversity within datasets persists. Therefore, our research aims to train the DNN model by using the StyleGAN model to expand the diversity of the data set, thereby improving the generalization of the DNN. Our approach targets specific attributes in the original image by controlling the latent space W of the StyleGAN, so that the generated image not only resembles the real-world image but also enriches the feature diversity of the dataset. Experimental results demonstrate that, in comparison to arbitrary data augmentation and class-balancing methods, the diversified datasets generated through this approach yield enhanced accuracy for VGG16 and LeNet models on the test set. Furthermore, our research employs the GD metric to quantitatively assess dataset diversity, thus substantiating the efficacy of the proposed methodology.