Nowadays, Convolutional Neural Nets (CNNs) have become the reference technology for many computer vision problems, including facial landmarks detection. Although CNNs are very robust, they still lack accuracy because they cannot enforce the estimated landmarks to represent a valid face shape. In this paper we investigate the use of a cascade of CNN regressors to make the set of estimated landmarks lie closer to a valid face shape. To this end, we introduce CRN, a facial landmarks detection algorithm based on a Cascade of Recombinator Networks. The proposed approach not only improves the baseline model, but also achieves state-of-the-art results in 300W, COFW and AFLW that are widely considered the most challenging public data sets.