High Quality Point Cloud Generation and Intuitive Text-Guided Editing Using Prior Latent Matrix

With the advent of affordable 3D printers and the development of virtual space technologies and services, the demand for 3D models has been increasing in recent years. Consequently, research on deep generative models for 3D shape modeling has been actively conducted. For a deep generative model to be useful as a 3D shape modeling tool, it must satisfy two requirements: high-quality shape generation and intuitive shape editing. For intuitive shape editing, we study a hybrid method of text guided editing based on a deep generative model and a user-friendly interface that allows for easy selection of editable regions. This enables easy and precise editing of specific shape regions. In order to enable detailed partial editing, a method equipped with prior latent matrix was adopted for the deep generative model. From experiments, we confirm high quality of generated 3D shape of the proposed generative model. We also evaluate flexibility of the proposed text-based editing of specific regions and efficiency of region selection methods.