Style transfer aims to render the content of a given image in the\ngraphical/artistic style of another image. The fundamental concept underlying\nNeuralStyle Transfer (NST) is to interpret style as a distribution in the\nfeature space of a Convolutional Neural Network, such that a desired style can\nbe achieved by matching its feature distribution. We show that most current\nimplementations of that concept have important theoretical and practical\nlimitations, as they only partially align the feature distributions. We propose\na novel approach that matches the distributions more precisely, thus\nreproducing the desired style more faithfully, while still being\ncomputationally efficient. Specifically, we adapt the dual form of Central\nMoment Discrepancy (CMD), as recently proposed for domain adaptation, to\nminimize the difference between the target style and the feature distribution\nof the output image. The dual interpretation of this metric explicitly matches\nall higher-order centralized moments and is therefore a natural extension of\nexisting NST methods that only take into account the first and second moments.\nOur experiments confirm that the strong theoretical properties also translate\nto visually better style transfer, and better disentangle style from semantic\nimage content.\n