LivingAvatars: Robust Head Reconstruction With Gaussian Lifecycle Management and Neural Detail Synthesis.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• 3D Gaussian Splatting (3DGS) enables high-fidelity avatar creation but faces representation and dynamic robustness bottlenecks.
• Existing methods struggle with modeling dynamic details and maintaining compact, robust representations, leading to information loss and point cloud inflation.

Purpose of the Study:

• To address the limitations of current 3D head avatar generation techniques.
• To introduce a novel framework, LivingAvatars, for creating robust and high-fidelity dynamic head avatars.

Main Methods:

• Developed a functionally decoupled Gaussian representation separating identity-stable base Gaussians from detail Gaussians.
• Implemented a lightweight neural decoder for synthesizing expression-dependent details.
• Introduced a Gaussian lifecycle management system with probabilistic selective pruning and temporary hiding to ensure compact and robust representations.

Main Results:

• LivingAvatars outperformed existing methods in novel-view synthesis and self-reenactment metrics.
• Achieved superior preservation of transient, expression-dependent signals in cross-identity reenactment, enhancing perceptual fidelity.
• Demonstrated robust and compact representation under dynamic occlusions.

Conclusions:

• Functionally decoupled Gaussian representations and lifecycle management are effective principles for robust neural head avatars.
• The proposed methods reliably reconstruct high-frequency details and improve overall realism in dynamic head avatar synthesis.
• LivingAvatars offers a significant advancement in creating compact, high-fidelity, and dynamically robust 3D head avatars.