This paper addresses the problem of learning based single image super-resolution. Previous research on this problem employs human user to provide a set of images that are similar to the target image as a reference. Then the super-resolution algorithm can learn from the provided reference images to predict the high resolution details for the target image. We propose a fully automatic scheme, which leverages the knowledge of the entire visual world and to query relevant references from the Internet. The proposed scheme is free of human supervision, and the performance compromise is small. We conduct experiments to show the effectiveness of the method.