Made-to-measure methods such as the parallel code NMAGIC have to solve underdetermined problems, where the number of constraints (photometric of kinematic observables) are usually much lower than the number of particle weights to define. A regularization method is therefore needed to make the method converge. Here we introduce a Moving Prior entropy Regularization method (MPR). The basic idea is to update the prior distribution needed by standard entropy regularization in parallel of the weight adaptation. The prior distribution is determined from the distribution of particles in phase-space. This allows one to construct smooth models from noisy data without erasing global phase-space gradients.