The article introduces Human Gaussian Splats (HUGS), a neural rendering framework designed for animatable human avatars and scenes. The framework trains on a short monocular video and learns to disentangle the static scene and a fully animatable human avatar within 30 minutes. HUGS uses 3D Gaussian Splatting (3DGS) and the SMPL body model to create detailed and accurate human models. The method also allows for the synthesis of novel human poses and views. HUGS achieves state-of-the-art rendering quality at a speed of 60 FPS, significantly faster than previous work.
Publication date: 29 Nov 2023
Project Page: https://github.com/apple/ml-hugs
Paper: https://arxiv.org/pdf/2311.17910
