A new preprint introduces GALA (Gaussian Animation via Linear Approximation), a way to cut repeated neural work in 3D Gaussian-avatar animation. In tests of three existing systems, it replaced much of the per-frame computation with reusable motion shapes and a coefficient predictor.
Gaussian avatars represent a person as many small 3D shapes that can be rendered quickly. But many systems use a neural network to recalculate each shape's properties, such as position, color and opacity, for every pose. GALA caches the neutral avatar, then represents movement as learned changes to those properties, called blendshapes. A coefficient predictor estimates how strongly to apply each change from the avatar's identity and pose; the system adds their weighted sum to the cached state.
The researchers built a separate motion basis for each pretrained system, then reused it across identities, including identities not used to build the basis. The method does not retrain the original avatar model. Its basis uses local groups of shapes and prioritizes changes expected to matter most in the rendered image.
Each timing used one animation step at a time, with eight CPU threads on an Intel Core Ultra 9 285K. The measured attribute-update step fell from 154 to 5.5 milliseconds for AGORA, from 234 to 4.4 milliseconds for FlexAvatar, and from 42.9 seconds to 6.1 milliseconds for DynaAvatar. The timings exclude rendering. DynaAvatar adds a 10-millisecond joint-based skinning step to both versions; including it makes the full animation step 42.9 seconds for the original and 16.1 milliseconds for GALA.
The authors’ phone-browser demos show one avatar at 60 frames per second on an iPhone 15, five heads at 40 to 50 frames per second, and five full-body avatars at 20 frames per second. The browser runs GALA’s predictor and blending steps rather than the original host network.
The motion is an approximation, not an exact copy: on DynaAvatar's 4D-Dress benchmark, peak signal-to-noise ratio, a measure of image similarity, was 0.64 decibels below the host model. The comparison is with the original system, not real people, and covers only these three selected models. The linked code repository says code is coming soon; broader device and model tests remain to be seen.
Sources
- Fazylov, Lefkimmiatis and Laptev, “One Basis to Animate Them All: Gaussian Blendshape Distillation for Real-Time Avatars,” arXiv:2610.02207v1 (October 1, 2026; preprint).
- Authors’ project page and mobile demos.
- Author-linked code repository (currently says code is coming soon).
Image: Figure 1 from Fazylov, Lefkimmiatis and Laptev, “One Basis to Animate Them All: Gaussian Blendshape Distillation for Real-Time Avatars,” arXiv:2610.02207v1. Used unchanged under CC BY 4.0.