Generative Modeling for 3D Shape and Scene Synthesis

by Katharina Schmid

3D shape generation is a rapidly growing area of research with broad relevance across domains such as computer graphics, virtual and augmented reality, and robotics. While recent generative models have achieved impressive results in producing realistic 2D images and videos, synthesizing high-quality 3D objects remains fundamentally more complex due to the added spatial dimension. This task becomes even more challenging when extending to full 3D scenes or dynamic objects, where relationships between multiple shapes and their temporal changes introduce additional layers of complexity.

Motivated by these challenges, my doctoral research focuses on developing novel 3D representations and generative modeling techniques aimed at advancing 3D asset synthesis. This includes exploring frameworks such as flow matching applied to 3D point clouds, alongside optimized strategies for matching spatial data to improve the stability and quality of generation. Moreover, I am interested in extending these approaches to richer and more expressive 3D representations, such as neural point clouds and Gaussian kernel-based models, with the goal of enabling more detailed, realistic, and controllable 3D content across a variety of applications.