Gaussian Head Avatar’s Venture Web page
Summary
Creating high-fidelity 3D head avatars has at all times been a analysis hotspot, however there stays a fantastic problem beneath light-weight sparse view setups.
On this paper, we suggest Gaussian Head Avatar represented by controllable 3D Gaussians for high-fidelity head avatar modeling.
We optimize the impartial 3D Gaussians and a totally realized MLP-based deformation subject to seize advanced expressions.
The 2 components profit one another, thereby our technique can mannequin fine-grained dynamic particulars whereas guaranteeing expression accuracy.
Moreover, we devise a well-designed geometry-guided initialization technique primarily based on implicit SDF and Deep Marching Tetrahedra for the soundness and convergence of the coaching process.
Experiments present our strategy outperforms different state-of-the-art sparse-view strategies, attaining extremely high-fidelity rendering high quality at 2K decision even beneath exaggerated expressions.
Fig 1.Gaussian head avatar achieves extremely high-fidelity picture synthesis with controllable expressions at 2K decision. The above reveals completely different views of the synthesized avatar, and the underside reveals completely different identities animated by the identical expression. 16 views are used through the coaching.