Yan Jiang

Facial expressions are one of the most intuitive and easily perceived signals of human emotions, and enhancing facial editing and generation can help models better understand these emotions. Yet, current generative editing methods face two major challenges: limited fine-grained controllability and privacy risks. This research adopts Action Units (AUS) as the foundation for controllable and trustworthy facial editing and generation. The work will: (1) improve AU detection and dependency modeling to enable single-image editing without reference images; (2) expand limited AU datasets via AU-consistent generative augmentation; (3) integrate multimodal cues such as infrared, text, and speech for context-aware editing; and (4) develop mechanisms to mark edited regions and prevent misuse of generated data. The resulting AU-driven framework aims to provide precise, robust, and ethically aligned facial editing, advancing trustworthy emotion understanding in real-world applications.