Relightable Gaussian Codec Avatars

Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, Giljoo Nam
Codec Avatars Lab, Meta

CVPR 2024 (Oral)
Paper arXiv

Relightable Gaussian Codec Avatars generalize to novel expressions, gaze, view points, and illuminations with global light transport and all-frequency reflections in real-time.

Abstract

In this work, we present Relightable Gaussian Codec Avatars, a method to build high-fidelity relightable head avatars that can be animated to generate novel expressions.

Our geometry model based on 3D Gaussians can capture 3D-consistent sub-millimeter details such as hair strands and pores on dynamic face sequences. To support diverse materials of human heads such as the eyes, skin, and hair in a unified manner, we present a novel relightable appearance model based on learnable radiance transfer. Together with global illumination-aware spherical harmonics for the diffuse components, we achieve real-time relighting with all-frequency reflections using spherical Gaussians. This appearance model can be efficiently relit in real-time under both point light and continuous illumination. We further improve the fidelity of eye reflections and enable explicit gaze control by introducing relightable explicit eye models.

Our method outperforms existing approaches without compromising real-time performance. We also demonstrate real-time relighting of avatars on a tethered consumer VR headset, showcasing the efficiency and fidelity of our avatars.

Video (w/ Audio)

Method

Relightable Gaussian Codec Avatars are conditioned with a latent expression code, gaze information, and a target view direction. The underlying geometry is parameterized by 3D Gaussians and can be efficiently rendered with the Gaussian Splatting technique. To enable efficient relighting with various illuminations, we present a novel learnable radiance transfer based on diffuse spherical harmonics and specular spherical Gaussians. Please refer to the paper for more details.

Disentangled Controls

Our model enables the disentangled control of expression, gaze, view, and lighting as demonstrated below.

Expression

Gaze

View

Lighting

You can try our interactive viewer with individual control of view/expression/light here.

Application

Interactive Relighting and Rendering in VR

Relightable Gaussian Codec Avatars can be rendered in real-time from any viewpoints of a VR headset. We also demonstrate interactive point light control as well as relighting in natural illumination.

Video-driven Animation

Relightable Gaussian Codec Avatars can be driven live from video streaming from head mounted cameras (HMC).

BibTeX


        @inproceedings{saito2024rgca,
          author = {Shunsuke Saito and Gabriel Schwartz and Tomas Simon and Junxuan Li and Giljoo Nam},
          title = {Relightable Gaussian Codec Avatars}, 
          booktitle = {CVPR},
          year = {2024},
        }