Recent Nonlocal Means filter (NLM) [1] is lack of mean conservation property, that is, image can have different mean before and after being processed. It is not suppose to happen when dealing with zero-mean noise. In this paper, a compensation procedure is proposed as post-process of NLM. The constraint of mean conservation is imposed on NLM through adding to the restored image of NLM with a compensation image. The compensation image is obtained via processing the method noise of NLM. Experiments show that the compensation procedure is able to significantly improve the performance of NLM.