Abstract. We propose a variational method for artifact-free video decompression that is capable of processing any MPEG-2 encoded movie. The method extracts, from a given MPEG-2 file, a set of admissible image sequences and minimizes an artifact-penalizing spatio-temporal regularization functional over this set, giving an optimal decompressed image sequence. For regularization, we use the infimal convolution of spatiotemporal Total Generalized Variation functionals (ICTGV). Numerical experiments on MPEG encoded files show that our approach significantly increases image quality compared to standard decompression.