Abstract
We present, GauHuman, a 3D human model with Gaussian Splatting for both fasttraining (1 ~ 2 minutes) and real-time rendering (up to 189 FPS), compared withexisting NeRF-based implicit representation modelling frameworks demandinghours of training and seconds of rendering per frame. Specifically, GauHumanencodes Gaussian Splatting in the canonical space and transforms 3D Gaussiansfrom canonical space to posed space with linear blend skinning (LBS), in whicheffective pose and LBS refinement modules are designed to learn fine details of3D humans under negligible computational cost. Moreover, to enable fastoptimization of GauHuman, we initialize and prune 3D Gaussians with 3D humanprior, while splitting/cloning via KL divergence guidance, along with a novelmerge operation for further speeding up. Extensive experiments on ZJU_Mocap andMonoCap datasets demonstrate that GauHuman achieves state-of-the-artperformance quantitatively and qualitatively with fast training and real-timerendering speed. Notably, without sacrificing rendering quality, GauHuman canfast model the 3D human performer with ~13k 3D Gaussians.