One Basis to Animate Them All: Gaussian Blendshape Distillation for Real-Time Avatars
GALA replaces expensive avatar animation decoding with a shared linear blendshape basis that can run at up to 60fps on mobile.
The paper says pretrained 3D Gaussian avatar models can be animated by approximating their outputs as a linear combination of identity-independent blendshapes. Its distillation method uses block-local PCA and a shallow MLP to predict coefficients, avoiding retraining the original models. The authors report tests on three avatar systems covering facial expressions and full bodies with clothing dynamics. They claim CPU animation cost drops by up to three orders of magnitude while preserving most rendering quality. ArXiv · AI/CL/LG's note
The paper says pretrained 3D Gaussian avatar models can be animated by approximating their outputs as a linear combination of identity-independent blendshapes. Its distillation method uses block-local PCA and a shallow MLP to predict coefficients, avoiding retraining the original models. The authors report tests on three avatar systems covering facial expressions and full bodies with clothing dynamics. They claim CPU animation cost drops by up to three orders of magnitude while preserving most rendering quality. ArXiv · AI/CL/LG's note
score 4