One Basis to Animate Them All: Gaussian Blendshape Distillation for Real-Time Avatars

TL;DR

The paper introduces GALA, a method that replaces expensive neural decoding in 3D Gaussian avatars with a lightweight linear blend of identity-independent blendshapes. It uses block-local PCA under a rendering-aware metric to construct a compact basis and trains a shallow MLP to predict blendshape coefficients, enabling real-time animation on mobile devices.

Why It Matters

Real-time 3D avatar animation on mobile hardware is now feasible without sacrificing visual quality, enabling applications in AR/VR, virtual meetings, and interactive media. The approach works across diverse avatar models and identities without retraining, reducing computational cost by up to 1,000x.

Key Findings

• Pretrained 3D Gaussian avatars exhibit a shared linear structure in their motion representation, enabling approximation via identity-independent blendshapes.
• Block-local PCA under a rendering-aware metric enables a compact, high-fidelity basis with reduced memory usage.
• A shallow MLP predictor for blendshape coefficients achieves real-time performance (up to 60fps) on mobile devices while preserving visual quality.