Super resolution (SR) in computer vision is an important task. In this paper, we compared several common used features in image super resolution of example-based algorithms. To combine features, we develop a cascade framework to solve the problems of deciding weights among features and improving computation efficiency. In the experimental results we can see the effectiveness of each independent or combined features. Finally, we modify the framework to have an adaptive threshold such that not only the computation load is much reduced but the modified framework is suitable to any query image as well as various image databases.