The color fidelity of displays such as LCD and PDP was commonly estimated by a series of color patches. However, no strong evidence showed that these color patches are well correlated to the real-world image characteristics. To minimize the gap between the color patches and the image characteristics, the present study first accumulated the statistics of different classes of image. Then, we chose several of them as reference images, manipulated their colors intentionally, and finally asked observers to compare the reference images with their color-perturbed counterparts. In the end, based on the classified image statistics, an empirical model was derived to predict the overall results of the visual image differences. In the model, the pixel frequencies (probability) of 125 color clusters summarized by a certain type of images were taken into account for predicting the displays' color fidelity on the type of images. The 125 color clusters were linked in a 5x5x5 cellular structure. If the color manipulation breaks the intera-relationship of the cellular structure, the reproduction would be unnatural. In contrast, if the manipulation shifts the colors without breaking the relationship, the reproduction still looks fine. We called the intera-relationship as Cellular Uniformity and integrated the measurement into our model. It could be an interesting idea for future development on image quality assessment.