Automatic Synthesis of Isotropic Textures on Surfaces from Sample Images

A fully automatic method of synthesizing isotropic textures on subdivision surfaces from sample images is presented. Both Gaussian and Laplacian pyramid representations of the sample texture are constructed. Texture synthesis proceeds coarse-tofine, by incrementally inverting the Laplacian pyramid to produce an initial guess and refining this guess using non-parametric sampling. The sampling procedure uses a nearest neighbor search while preserving first-order statistics. The resulting texture is generated directly on the subdivision surface. Within the domain of isotropic textures, the proposed method offers improvements in faithful reproduction of a sample's appearance over a wide range of scales. The method can also be used to produce isotropic variants of anisotropic textures. Finally, while the sampling procedure we describe is not amenable to standard methods for nearest neighbor search in high dimensional spaces, an acceleration method is proposed that uses an eigenvector transform and a set of