One Basis to Animate Them All: Gaussian Blendshape Distillation for Real-Time Avatars
GALA approximates neural 3D avatar animation with linear blendshapes, cutting CPU cost up to 1,000x and reaching 60fps on mobile.
The paper introduces GALA (Gaussian Animation via Linear Approximation), which distills pretrained 3D Gaussian avatar models so per-frame neural decoding is replaced by a shallow MLP coefficient predictor and a linear blend of identity-independent blendshapes. The basis is built with block-local PCA under a rendering-aware metric and a memory budget, and the method applies without retraining the original models. Across three avatar models for facial expressions and clothed full bodies, GALA generalizes to held-out identities, cuts CPU animation cost by up to three orders of magnitude, and reaches about 60 frames per second on mobile devices while preserving most rendering quality.
- GALA replaces heavy per-frame neural decoding with a shallow MLP and linear blendshapes.
- Block-local PCA builds a compact basis under a rendering-aware metric and memory budget.
- Tested on three avatar models for faces and clothed full-body motion.
- CPU animation cost drops by up to 1,000x, with about 60fps on mobile devices.
From the original article
1 min read at arxiv.org3D Gaussian avatars support fast rendering, however, their real-time animation is often challenged by the costly neural inference. We address this bottleneck and show that the animation of pretrained avatar models can be closely approximated by a linear combination of identity-independent blendshapes.
By Ramazan Fazylov · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.