Curvelet is a multiscale and multidirectional image transformation tool, which can efficiently overcome the redundancy of wavelet in expressing the singular feature along curves of the image, and can obtain a sparse feature representation. Moreover, based on the consideration that high-dimensional image may exist in lower dimensional manifolds, manifold learning is performed on the Curvelet features so as to find low-dimensional structures, which is used for face recognition. Experiments show that the Curvelet features further processed by LLE show better clustering ability than the LLE. Compared with the already existing Gabor-based manifold learning, Curvelet-based manifold learning perform better under both facial expression and illumination changes, and either case sees valuable improvements. Experiments in the Essex expression and Yale B lighting face databases prove this point.