In this paper, we introduce a novel model based on Variational Auto-Encoder (VAE) that is able to find subclasses of categories and generate new samples with fidelity to the subclass in an unsupervised way. In generating characters from historical documents, this model helps generated characters to avoid ambiguity in the case where there are multiple writing styles of one character without being labeled. With this model we augment historical Japanese document dataset to make it more balanced. The model is trained in two steps. In the first step, the model learns the data distribution and learns to map character images into basic shape vectors. In the second step, the model learns to generate new samples conditioned on the basic shape vectors. The generated samples are more robust against intra-class multi-modality. With the usage of augmented dataset, the recognition rate is improved. Ablation study is performed to evaluate the effectiveness of data augmentation.