A probabilistic grammar for the grouping and labeling of parts and objects, when taken together with pose and part-dependent appearance models, constitutes a generative scene model and a Bayesian framework for image analysis. To the extent that the generative model generates features, as opposed to pixel intensities, the inverse or posterior distribution on interpretations given images is based on incomplete information; feature vectors are generally insufficient to recover the original intensities. I will argue for fully generative scene models, meaning models that in principle generate actual digital pictures. I will outline an approach to the construction of fully generative models through an extension of context-sensitive grammars and a re-formulation of the popular template models for image fragments. Mostly I will focus on the problem of learning template models from image data. Since the model is fully specified (generative), at the pixel level, the templates can be learned by maximum likelihood. A training set of eyes, for example, yields an ensemble of left and right eyes, of familiar and natural character, but not actually coming from any particular individuals in the training set. The upshot is a mixture distribution on image patches, consisting of a set of templates and a set of conditional patch distributions - one for each template. One way to test the model is to examine samples. I will show how to sample from the mixture distribution and I will show sample sets of eyes, mouths, and generic background. Another way to test the model is to use it for detection, recognition, or classification. I will show the results of a test on ethnic classification based on the eye region of faces.