Reconstruction of a high-resolution face image, from a low-resolution observation based on a set of high- and low-resolution training image pairs, is an important problem for optical engineering applications. In this paper, we study this facial superresolution problem and propose a novel locally nonlinear transformation based approach. Multiple locally nonlinear transformation are utilized to approximate the global nonlinear connections between low resolution (LR)/high resolution (HR) images. LR/HR images are initially divided into multiple pairs of patches with the corresponding position information. As facial images are highly structured, patches at the same position spanned a subspace. Since the curse of dimensionality is avoided in these subspaces (patches in the same position), the Euclidean distance can express the intrinsic “radial” between samples in the same subspace. Therefore, multiple radial basis functions are utilized to approximate the nonlinear mapping between LR/HR pairs at each position from training examples. The proposed locally nonlinear transformation (LNT)-based reconstruction is achieved by applying the learned nonlinear transformation to each position patch of an LR input. The final SR results are obtained by refining the LNT reconstruction by the projection onto a convex sets algorithm using the consistency constraint. Extensive experiments on benchmark databases and real world images validate the superiority of the proposed method.