Bayesian analysis is a popular subspace based face recognition method. It casts the face recognition task into a binary classification problem with each of the two classes, intrapersonal variation and extrapersonal variation, modeled as a Gaussian distribution. However, with the existence of significant transformations, such as large illumination and pose changes, the intrapersonal facial variation cannot be modeled as a single Gaussian distribution, and the global linear subspace often fails to deliver good performance on the complex non-convex data set. We extend the Bayesian face recognition into Gaussian mixture models. The complex intrapersonal variation manifold is learnt by a set of local linear intrapersonal subspaces and thus can be effectively reduced. The effectiveness of the novel method is demonstrated by experiments on the data set from AR face database containing 2340 face images.