An investigation of portrait relighting trajectories in the latent space of deep generative models

Deep learning models are gaining more and more success because of their great results and the wide application fields. An instance of deep learning models are the Deep Generative Models (DGMs) whose aim is to learn the probability distribution from independent and identically distributed samples and create new samples whose distribution is similar to the starting one. The goal of this dissertation is to analyze the DGMs and the manipulation of them to force the generation of samples with specific characteristics. In particular, it was analyzed how generated portrait images could be modified depending on the light position. The portrait relighting problem is challenging because human observers are sensitive to the subtleties of facial appearance and have a low tolerance to errors in processed face images. To understand this task in depth, various methods were inspected going from pure physical models to pure deep learning approaches as well as mixed approaches. The work here exposed uses physical approximations on lighting to transfer the task of relighting in the field of latent space manipulation. The results obtained can be considered satisfying for the complex reflections and shadows on relighted images that produce a natural and pleasing effect.