The main cluster algorithm in image clustering especially visual dictionary construction is still k-Means currently. The limitation of k-Means seriously deteriorates its feasibility for large image set, so we propose an Ensemble based Locality Sensitive Clustering method, which is based on distance and separability preservation property of random projection and the rationale of Exact Euclidean Locality Sensitive Hashing. It first determines the number of clusters of dataset, and then generates the multiple clustering resolutions, at last, applies cluster ensemble methods to get final partition. This method can make use of the advantage of Locality Sensitive Hashing, and improve clustering accuracy by cluster ensemble. The experiments showed that it achieved high cluster accuracy. In addition to its advantage of fast running speed, low cost and dynamic clustering, Ensemble based Locality Sensitive Clustering is a promising method for image cluster.