Illumination harmonization is an important problem for high-quality image composite. Given the source image and the target background, it aims to transform the foreground appearance that it looks in the same lighting condition as the target background. Because the ground-truth composite image is difficult to get, previous works can use only synthetic datasets, which however, provide with only artificially adjusted and limited inputs. In this paper we contribute to this problem in two aspects: 1) We introduce a semi-automatic approach to capture the ground-truth composite in real scenes, and then create a dataset that enables faithful evaluation of image harmonization methods. 2) We propose a simple yet effective harmonization method, namely the Gray Mean Scale (GMS), which models the foreground appearance transformation as channel-wise scales, and estimates the scales based on gray pixels of the source and the target background images. In experiments we evaluated the proposed method and compared it with previous methods, using both our dataset and previous synthetic datasets. A new benchmark thus is established for illumination harmonization in real environments.