We address the problem of determining the \emph{optimal model complexity} for shape modeling. This complexity is a compromise between model specificity and generality. We show that the error of a model can be split into two components, the model error and the fitting error, of which the first one can be used to optimize the model complexity based on the specific application. This strategy improves over traditional approaches, where the model complexity is only determined by vague heuristics or trial-and-error. A method for the determination of optimal active shape models is proposed and its efficiency is validated in several experiments. Furthermore, this method gives an indication on the range of valid shape parameters and on whether or not an increased number of training data will reduce the number of shape parameters further.