Developing Level 3 or higher autonomous vehicles requires the ability to follow human traffic controllers in situations where regular traffic signals are unavailable, such as during construction. However, detecting human traffic controllers at construction sites is challenging due to the lack of dedicated datasets and variations in their appearance. This paper proposes a method for detecting human traffic controllers by generating synthetic images with diffusion models. We introduce a color-boosting technique to enhance image diversity and employ a cut-and-paste mechanism for seamless integration into realistic road scenes. We generate 19,840 synthetic images, combined with 600 real-world images, to train a YOLOv7 model. The trained model achieves an