Robin Pourtaud

In recent years, the adoption of 3D models has increased across various fields, includ- ing entertainment, medical imaging, and virtual reality. Concurrently, generative AI mod- els have demonstrated significant success in generating diverse media types, such as text, images, and videos. Inspired by the principles of DALL-E 2 and Stable Diffusion models, which generate high-quality 2D images from textual inputs, this thesis aims to investigate the feasibility of extending these principles to generate 3D mesh structures from graph repre- sentations. To achieve this, we explore suitable Graph Neural Networks (GNNs) for graph embedding and adapt diffusion methods for 3D mesh data. Our proposed approach con- tributes to the understanding of generative models for 3D mesh structures and serves as a stepping stone towards generating more complex networks, such as knowledge graphs, which have wide-ranging applications in semantic web services, information extraction, and knowledge management.