Fashion companies have always been facing a critical issue to design products that fit consumers’ needs. On one hand, fashion industries continually reinventing itself. On the other hand, shoppers’ preference is changing from time to time. In this work, we make use of machine learning and computer vision technologies to automatically design new “must-have” fashion products with popular styles discovered from fashion product images and historical transaction data. Products in each discovered style share similar visual attributes and popularity. The visual-based fashion attributes are learned from fashion product images via a deep convolutional neural network (CNN). Fusing together with popularity attributes extracted from transaction data, a set of styles is discovered by Nonnegative matrix factorization(NMF). Eventually, new fashion products are generated from the discovered styles by Variational Autoencoder (VAE). The result shows that our method can successfully generate combinations of interpretable elements from different popular fashion products. We believe this work has the potential to be applied in the fashion industry to help to keep reasonable stocks of goods and capture most profits.