An efficient and accurate face modeling based on Garbor wavelet network (GWN) is introduced. GWN is very useful because it can create sparse representation of object for fast matching and it is a biologically motivated feature. However the beauty of GWN is offset by inefficient time-consuming modeling of objects. Therefore an efficient and accurate modeling method is needed for GWN. GWN prior for faces is learned from University of Essex Faces94 database using Levenberg-Marquardt optimization method and K-means clustering. Then, guided by GWN prior, the new modeling process efficiently achieves better modeling. To evaluate the quality of modeling, sum of squared difference is used and tracking algorithm is applied to the model as a verification of our proposed method.