This thesis introduces multiple innovative frameworks for 3D generation, with an emphasis on not just creating unique 3D shapes in varied representations but also introducing different mechanisms to control the generation processes while demonstrating the possibility of learning 3D generative models on a large scale.
- we propose a novel framework for directly generating 3D meshes, which decomposes the generative task into two distinct sub-tasks: topology formation and shape deformation.
- we present a novel generative framework for creating high-resolution implicit functions.
- we extend the wavelet-based generation framework for shape inversion task.
- we delve into large-scale 3D generation, proposing a new 3D generative model trained on an extensive dataset of 10 million publicly available shapes.