Published on:
8 April 2024
Primary Category:
Computer Vision and Pattern Recognition
Paper Authors:
Heyuan Li,
Ce Chen,
Tianhao Shi,
Yuda Qiu,
Sizhe An,
Guanying Chen,
Xiaoguang Han
Proposes SphereHead method for full 3D head synthesis
Uses spherical tri-plane representation to avoid feature entanglement
Introduces view-image consistency loss for better image-view alignment
Generates artifact-free 360 degree head views
Publicly releases code and new multi-view head image dataset
Stable 3D full-head image synthesis
This paper proposes a new method called SphereHead to address challenges in synthesizing realistic 3D renderings of full human heads from all angles. It introduces a spherical tri-plane feature representation that avoids artifacts caused by feature entanglement in prior methods. Additionally, a new view-image consistency loss improves results by enforcing better alignment of rendered images with their corresponding viewing angles. Experiments demonstrate SphereHead's ability to generate high-quality, artifact-free 360 degree head views.
No comments yet, be the first to start the conversation...
Sign up to comment on this paper