The article introduces Gaussian Shell Maps (GSMs), a new framework for efficient 3D human generation. It connects state-of-the-art generator network architectures with 3D Gaussian rendering primitives using a multi-shell based scaffold. A CNN generates a 3D texture stack with features mapped to the shells, representing various versions of a digital human template. This system allows for the creation of high-quality, multi-view consistent renderings at a high resolution, bypassing the need for slow and inconsistent 2D upsamplers. The GSMs have been demonstrated to successfully generate 3D humans when trained on single-view datasets.
Publication date: 29 Nov 2023
Project Page: rameenabdal.github.io/GaussianShellMaps
Paper: https://arxiv.org/pdf/2311.17857
