EAvatar: Expression-Aware Head Avatar Reconstruction with Generative Geometry Priors

Code

Shikun Zhang1, Cunjian Chen1, Yiqun Wang2*, Qiuhong Ke1, Yong Li2*

1Department of Data Science and AI, Monash University, Melbourne, VIC 3800, Australia

2College of Computer Science, Chongqing University, Chongqing 401331, China

* Corresponding authors.

Network Architecture Overview

The overview of the EAvatar rendering and reconstruction. We first learn an implicit SDF-based geometry and extract the surface via DMTet. A high-quality mesh from a large-scale pretrained model is used as a generative prior to stabilize initialization and guide accurate shape construction. In the second stage, we build upon the predicted mesh to further refine the dynamic Gaussian representation. A controllable Gaussian mechanism and a splitting strategy are introduced to improve expression-driven deformation and local detail.

Abstract

We propose a novel expression-aware 3D head avatar reconstruction framework with a controllable Gaussian mechanism that enables expression-driven animation and accurate reproduction of fine-grained expressive details.

We design a Gaussian splitting strategy to enhance the geometric expressiveness in high-deformation regions. We introduce a structure-aware geometry modeling module guided by generative priors from a large-scale generative model, which improves early-stage training stability and ensures globally consistent geometry

Our method is evaluated on multiple expression-driven benchmarks. The results demonstrate superior performance in terms of expression reconstruction accuracy, detail preservation, and identity consistency, showing strong generalization and practical value.

Comparisons on cross-identity reenactment task

Qualitative comparisons of different methods on cross-identity reenactment task. From left to right: NeRFace, HAvatar, GHA and Ours. Our method accurately reproduces the migration of expressions while synthesizing high-quality images.

Comparisons on self-reenactment task

Qualitative comparisons of different methods on self-reenactment task. From left to right: NeRFace, HAvatar, GHA and Ours. Our method can reconstruct details like eyes, teeth, etc. with high quality

Ablation study on Control Mechanism

Ablation study on Control&split Mechanism