In this paper, we propose a facial modeling system using our novel multilevel adaption procedure. We model a template head model with physical structure to synthesize the various facial expression. To make the template model suitable for different facial geometry, landmarks compliant to MPEG-4 Facial Definition Parameters (FDP) are labeled on this model. Then another personalized 3D head with different geometrical information is labeled with a subset of these feature points. An iterative multilevel training process is performed to get a transformation function from the template model to the personalized model. We use Compactly Supported Radial Basis Function (CSRBF) to assure the locality detail of the adapted model. A curvature based searching scheme is also used to find the feature points in the iterative process.