Although CSVD can eliminate the SVD-caused intrinsic defect that the basic spaces of reconstructed image are disagreed, the features of image classification are not impaired. Because of the image vectorization, the face recognition algorithm based on NMF will cause the lost of structure information and takes more memory. Although 2DNMF avoid these shortcomings caused by NMF, its own defect that the slow iterative convergence speed and the long training time will appear along with the increase of training samples. By combining the advantages and disadvantages of CSVD and 2DNMF, this paper advances the Joint CSVD-2DNMF face recognition algorithm. Experimental results from ORL face image database by using Matlab show that the efficiency of this advanced fusion method can shorten training time and improve recognition rates effectively.