Evaluating Text-to-Image Generative Models: An Empirical Study on Human Image Synthesis