Face photo sketch synthesis has got some researchers' attention in recent years because of its potential applications in digital entertainment and law enforcement. Some patches based methods have been proposed to solve this problem. These methods usually focus more on how to get a sketch patch for a given photo patch than how to blend these generated patches. However, without appropriately blending method, some jagged parts and mottled points will appear in the entire face sketch. In order to get a smoother sketch, we propose a new method to reduce such jagged parts and mottled points. In our system, we resort to an existed method, which is Markov Random Fields (MRF), to train a crude face sketch firstly. Then this crude sketch face sketch will be divided into some larger patches again and retrained by Non-Negative Matrix Factorization (NMF). At last, we use Multiresolution Spline and a blend trick named full-coverage trick to blend these retrained patches. The experiment results show that compared with some previous method, we can get a smoother face sketch.