A method taking multi-samples as sub-modes an d grouping face modes into partial intersection ones was proposed to reduce comput ation and improve system extension property. In combination, the sum rule based on Bayesian theory was used. The face recognition experiments with the ORL and A R face databases showed that the eigenface algorithm using multi-samples reache d a high recognition rate and a reasonable time cost. Grouping face modes makes training become a distributed computation job which reduces time cost for traini ng and brings the convenience of system extension when new face modes are to be added.