A method for example-based face hallucination uses manifold learning to project a plurality of training images in a training database and an input low resolution (LR) face image into a same manifold domain, then iteratively refines the reconstruction basis by selecting a training set having k projected training images which best match the parts of the projected LR face image, where k≦N and N is the number of projected training images. Through the best-match training set, a set of prototype faces are learned, and the set of prototype faces are used as the reconstruction basis to reconstruct a high resolution face image for the input LR face image.