This paper describes a method for colorimetric color reproduction on a dye sublimation printer by means of neural networks. A multilayer feed forward network is regarded as a nonlinear transformer for color coordinate transformation between the printer coordinates and the color stimulus values. The network is trained to learn a mapping to determine the required CMY (RGB) values of printer primaries for producing a given XYZ color stimulus. We adopt the Back-Propagation learning rule far the training. The mapping is then realized in a simple network architecture in which nonlinear units are linked in parallel and in layers. The measured data of many color patches are used for training the network and testing the mapping accuracy. The accuracy is evaluated on the CIE-L*a*b* color difference between the reproduced color from the network output and the original color. We determine an effective network method based on experiments under different conditions.