IA-3DGS Identity-Anchored Dynamic Gaussian Splatting for Long-Term Facial Detail Preservation

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Dynamic 3D Gaussian Splatting (3DGS) offers efficient rendering for facial avatars.
• Long monocular sequences often lead to cumulative geometric drift and loss of identity details with standard 3DGS.
• Maintaining identity consistency over extended periods remains a challenge in dynamic avatar rendering.

Purpose of the Study:

• To develop an identity-anchored dynamic Gaussian framework (IA-3DGS) for robust long-term facial animation.
• To address geometric drift and identity degradation in monocular facial sequences.
• To enhance the realism and temporal consistency of dynamic facial avatars.

Main Methods:

• Employed a 3D morphable model-guided semantic initialization to associate Gaussian primitives with facial regions.
• Introduced a region-weighted Jacobian rigidity regularizer to control deformation in quasi-rigid and expression-sensitive facial areas.
• Implemented a cyclic memory correction mechanism with exponential moving average smoothing for temporal stability.

Main Results:

• Achieved high performance on standard sequences (PSNR: 31.85 dB, SSIM: 0.942, LPIPS: 0.038).
• Demonstrated superior long-term identity consistency at frame 5000 (L-LPIPS: 0.046, landmark error: 1.3 mm) compared to baseline.
• Rendered high-resolution (1920x1080) video at ~48 FPS on a single GPU.

Conclusions:

• The proposed semantic, geometric, and temporal constraints effectively improve identity consistency in long-sequence facial animation.
• IA-3DGS provides a robust solution for generating high-fidelity dynamic facial avatars from monocular video.
• The framework shows significant potential for applications requiring long-term, identity-preserving facial rendering.