For multi-view face alignment, we have to deal with two major problems: 1. the problem of multi-modality caused by diverse shape variation when the view changes dramatically; 2. the varying number of feature points caused by self-occlusion. Previous works have used nonlinear models or view based methods for multi-view face alignment. However, they either assume all feature points are visible or apply a set of discrete models separately without a uniform criterion. In this paper, we propose a unified framework to solve the problem of multi-view face alignment, in which both the multi-modality and variable feature points are modeled by a Bayesian mixture model. We first develop a mixture model to describe the shape distribution and the feature point visibility, and then use an efficient EM algorithm to estimate the model parameters and the regularized shape. We use a set of experiments on several datasets to demonstrate the improvement of our method over traditional methods.