Treffer: Generic face attribute manipulation

Ph.D. ; Digital face manipulation has become a popular and fascinating way to touch images with the prevalence of smartphones and social networks. With a wide variety of user preferences, facial expressions, and accessories, a general and flexible model is necessary to accommodate different types of facial editing. In this thesis, we provide an in-depth analysis of this task, and propose four frameworks that address problems of this task. ; First we propose a model, called Facelet-Bank, to support fast inference, edit-effect control, and quick partial-model update. Compared with conventional task-specific face manipulation methods, this model can deal with different attributes. Besides, it does not follow the conventional solutions that rely on adversarial training. Alternatively, by leveraging pseudo labels and conducting latent space interpolation, the model can be trained stably and robustly, leading to high quality and resolution results. In addition, this framework is highly flexible – each operation is related to only a few computed convolutional layers. Experimental results show that this framework can handle a wide range of expressions, accessories, and makeup effects. It produces high-resolution and high-quality results in fast speed. ; We further improve the flexibility of the Facelet-Bank model with Semantic Component Decomposition. Semantic Component Decomposition decomposes a facial attribute into multiple semantic components, each corresponds to a specific face region. This not only allows for user control of edit strength on different parts based on their preference, but also makes it effective to remove unwanted edit effect. Further, each semantic component is composed of two fundamental elements, which determine the edit effect and region respectively. This property provides fine interactive control. As shown in experiments, this model not only produces high-quality results, but also allows effective user interaction. ; In the third work, we propose another interpolation-based framework called ...