Accurate discrimination of images features is a main success factor towards efficient content-based image retrieval systems. These features can be extracted using local and/or global descriptors. Researchers efforts showed that, hybrid descriptors reported superior results compared to methods that use single descriptor, where hybridization certainly complements benefits from different perspectives. Genetic Algorithm (GA) is a heuristic computational intelligence approach that can be used to achieve the optimal satisfactory user image retrieval requests. In this paper, a new hybrid efficient and effective evolutionary retrieval approach (CBIR-GAF) based on late fusion of four global descriptors is proposed. Each descriptor produces a list of retrieved similar images to user query image and if these lists are merged correctly by late fusion, the results are improved. Thus, GA occurs to assign different weights to each retrieved image while merging, and then it optimizes these weights with a suitable fitness function to select optimum heterogeneous retrieved images. The proposed approach is evaluated on two benchmark datasets (Inria Holidays and Oxford5k), and reported a promising results where it enhanced the average accuracy in comparison of literature techniques.