In production environments, facial rigging is commonly done by either geometric deformations or blendshapes. Geometric deformations are driven by simulated muscle actions, which are loosely based upon the dynamics of facial tissue [Magnenat-Thalmann et al. 1988]. Blendshapes interpolate a large number of sculpted shapes [Bergeron and Lachapelle 1985]. The former approach is intuitive, yet slow and less precise. The latter is fast, yet memory intensive and sensitive to model changes. Conventional implementations of both approaches are difficult to generalize in order to build rigs quickly and retarget animation efficiently.