On May 25, 2026, the 89th Zhi Xin Forum was held in Lecture Hall 117 of the Zhi Xin Building. Professor Chia-Wen Lin from Tsing Hua University (NTHU), Taiwan, China, was invited to deliver a lecture titled “Physics-Informed Generative Image Restoration: From Data to Models.”
In this talk, Professor Lin systematically introduced how to leverage physics-based models to enhance image restoration performance: first, using physics-based models to generate synthetic training data with realistic degradation distributions to effectively augment datasets; second, employing physics-informed priors to guide the reverse process of diffusion models, thereby improving restoration fidelity and robustness while preserving their generative capability. Furthermore, Professor Lin presented representative experimental results of this approach on image dehazing and deblurring tasks. The results demonstrate that the diffusion restoration model incorporating physics-informed priors outperforms traditional supervised learning methods in both subjective visual quality and objective evaluation metrics, especially exhibiting stronger generalization ability when handling unseen scenes.
After the report, Professor Lin had a warm exchange and discussion with the attending faculty and students on topics such as the integration strategies of physics-based models and data-driven models, as well as the sampling efficiency of diffusion models. Drawing from his own experience, he also encouraged students to actively broaden their horizons, explore diligently, and discover and solve new scientific problems. This report further expanded the horizons of our university's faculty and students and deepened their understanding and knowledge of image restoration algorithms.
