Generative segmentation methods such as the Active Appearance Models (AAM) establish dense correspondences by modelling variation of shape and pixel intensities. Alas, for 3D and high-resolution 2D images typical in medical imaging, this approach is rendered infeasible due to excessive storage and computational requirements. This paper extends the previous work of Wolstenholme and Taylor where Haar wavelet coefficient subsets were modelled rather than pixel intensities. In addition to a detailed review of the method and a discussion of the integration into an AAM-framework, we demonstrate that the more recent bi-orthogonal CDF 9-7 wavelet offers advantages over the traditional Haar wavelet in terms of synthesis quality and accuracy. Further, we demonstrate that the inherent frequency separation in wavelets allows for simple band-pass filtering, e.g. edge-emphasis. Experiments using Haar and CDF 9-7 wavelets on face images have shown that segmentation accuracy degrades gracefully with increasing compression ratio. Further, a proposed weighting scheme emphasizing edges was shown to be significantly more accurate at compression ratio 1:1, than a conventional AAM. At higher compression ratios the scheme offered both a decrease in complexity and an increase in segmentation accuracy.