A practical neural network model was designed to realize the color space conversion of digital photofinishing. The sampling, network structure and training process were introduced respectively. But in actual training, the networks fall into local minimum in all probability. To solve this problem, evolutionary programming (EP) algorithm was applied and the learning rate was adaptively adjusted. In the experiment, the performance of network was compared with pre-optimizing. Then the color space conversion was evaluated by the simulation error of samples from the point of color difference.