This paper deals with the parameter choice for the NL-Means algorithm. Starting with basic computations on toy models, we study the bias-variance trade-off of this filter using a simple notion of regularity in the patch space. We show that this regularity is necessarily local and so should be the parameters of the filter. Relying on Stein's Unbiased Risk Estimate, we then propose a way to locally set these parameters, and we compare this method with the Non-Local Means with optimal global parameter.