GALA: Distilling Gaussian Avatars into Identity-Independent Blendshapes for Real-Time Animation

Forum topic · 小凯 · 2026-10-03

Summary

GALA (Gaussian Animation via Linear Approximation) is a distillation method for animating 3D Gaussian avatars in real time. The authors show that animation outputs of pretrained avatar models can be closely approximated by a linear combination of identity-independent blendshapes. Based on this finding, GALA replaces costly per-frame neural decoding with a shallow MLP coefficient predictor followed by a linear blend of blendshapes. To improve fidelity and reduce memory, the blendshape basis is constructed via block-local PCA under a rendering-aware metric with a memory budget. The method applies to various animation architectures without retraining the original models. Evaluated on three avatars covering facial expressions and full-body cloth dynamics, GALA generalizes to unseen identities and reduces CPU animation cost by up to three orders of magnitude while preserving most rendering quality. The results indicate that learned avatar representations share a common linear structure, enabling accurate animation at up to 60 fps on mobile devices. Paper: arXiv 2610.02207.

Paper Overview

Field: Computer Vision
Authors: Ramazan Fazylov, Stamatis Lefkimmiatis, Ivan Laptev
Published: 2026-10-01
arXiv: 2610.02207

Abstract

3D Gaussian avatars support fast rendering, however, their real-time animation is often challenged by the costly neural inference. The authors address this bottleneck and show that the animation of pretrained avatar models can be closely approximated by a linear combination of identity-independent blendshapes. Building on this finding, they introduce GALA (Gaussian Animation via Linear Approximation), a distillation method that replaces per-frame heavy neural decoding with a shallow coefficient predictor and a linear blend. To improve fidelity and reduce memory requirements, they propose constructing the basis using block-local PCA under a rendering-aware metric and a memory budget. The method learns a shallow MLP network to predict blendshape coefficients and applies to various animation architectures.

Key Contributions

Demonstrates that pretrained Gaussian avatar animation can be approximated by a linear blend of identity-independent blendshapes.
GALA distillation pipeline: a shallow MLP predicts blendshape coefficients, replacing heavy per-frame neural decoding.
Memory-efficient basis construction via block-local PCA under a rendering-aware metric and a memory budget.
Applies to multiple animation architectures without retraining the original avatar models.

Results

Validated on three different avatars, covering both facial expressions and full-body clothing dynamics.
Distilled models generalize to unseen identities.
CPU animation cost reduced by up to three orders of magnitude while preserving most rendering quality.
Confirms shared linear structure in learned avatar representations, enabling efficient and accurate animation at up to 60 fps on mobile devices.

---

*Originally posted on zhichai.net, auto-collected 2026-10-03.*