In this paper we propose a method to register a pair of images unseen to the original dataset based on a generative manifold model. The basic premise of this approach is to design an image distance metric using a weighted sum of similarity and smoothness terms derived from a diffeomorphic registration of pairwise images. A refined image distance matrix based on this metric can be adopted as an input for nonlinear dimensionality reduction of the dataset, and the learned manifold can be approximated to simultaneously reflect the variations of appearance and anatomical shape. The generative manifold model that combines the image distance measurement and the manifold learning technique is used to estimate the geodesic path via the unseen pair for composition of the final deformation field. The experimental result of a set of real 3D mouse brain volumes demonstrates that the estimated manifold coordinates appropriately reflect the trend in the original dataset and that the registration of unseen images using the shortest path inferred from the generative manifold model improves the result against the direct registration.