Modern image classifiers are often suffering over-fitting problems because of the insufficient number of images in the dataset. Data augmentation is a strategy to increase the number of training samples. However, recent data augmentation methods are designed manually and cannot generate real-like images. Some neural network-based image generation methods such as GAN and VAE can also be used for data augmentation, but they are usually applied to unbalanced datasets. Since the generated images cannot be guaranteed to be from the same label, using them to extend a balanced dataset may lead to decreasing the accuracy of the classifier. In this paper, we propose an image transfer network to produce images that automatically adapt to a specific dataset and classifier. The image transfer network will search for the output images which can maximize the validation accuracy and help the classifier to overcome the over-fitting problems. Through the experiments, our method achieves high accuracy on CIFAR-10 and CIFAR-100 datasets. Moreover, since it could combine with other data augmentation methods, we show that using our method can push the state-of-the-art results furthermore.