Single image super resolution is one of the most important topic in computer vision and image processing research, many convolutional neural networks (CNN) based super resolution algorithms were proposed and achieved advanced performance, especially in recovering image details, in which PixelCNN is the most representative one. However, due to the intensive computation requirement of PixelCNN model, running time remains a major challenge, which limited its wider application. In this paper, several modifications are proposed to improve PixelCNN based recursive super resolution model. First, a discrete logistic mixture likelihood is adopted, then a cache structure for generating process is proposed, with these modifications, numerous redundant computations are removed without loss of accuracy. Finally, a partial generating network is proposed for higher resolution generation. Experiments on CelebA dataset demonstrate the effectiveness the superiority of the proposed method.