Automatic image colorization is an old problem in image processing that has regained interest in the recent years with the emergence of deep-learning approaches with dramatic results. A careful examination shows that these methods often suffer from the so-called "color halos" or "color bleeding" effect: some colors are not well localized and may cross shape edges. This phenomenon is caused by the non-alignment of edges in the luminance and chrominance maps. We address this problem by regularizing the output of an efficient image colorization method with deep image prior and coupled total variation.