Addressing issues such as unnatural editing results and the inability to effectively simulate occluded content in images, an algorithm is proposed for image editing by discovering the semantic direction of local image regions. First, the masks of the editing region and the remaining region are obtained, and their Jacobian matrices are computed using a pre-trained StyleGAN2 model and the corresponding masks. Then, based on the principle that the editing region undergoes the most significant changes while the remaining region experiences minimal alterations, the semantic discovery of the local region is formulated as a two-layer optimization problem. Finally, leveraging optimal transport theory, an algorithm is proposed to determine the semantic direction by utilizing the correlation between the editing and remaining regions. This approach not only provides a closed-form solution but also ensures high efficiency. Experimental results demonstrate that, compared to mainstream local image editing algorithms such as LowRankGAN and ReSeFa, the proposed algorithm effectively simulates occluded content and achieves natural image editing, particularly when modifying facial features such as the mouth and eyes. While maintaining the quality of the edited image, it also enables precise local region editing, confirming the algorithm’s controllability and effectiveness.