Based on muscle control items, a uniform model for facial action was constructed, which consisted of time series and the corresponding vector series. With the top-down motion planning, the facial action could be decomposed to motion unit by time, type of motion, singular point and control item. To solve the inconsistent problem of adjacent motion units, a joining algorithm was designed according to their relationship. Nested script of facial action was defined to implement the complex facial action. Experiments have shown that the methods can implement continuous and complicated facial action realistically and effectively.