3D Gaussian Blendshapes for Head Avatar Animation

How to read this

Category

Method: photorealistic head avatar (Gaussian-splatting blendshapes)

Contributions

• Represents a neutral head model and a group of expression blendshapes entirely as 3D Gaussians, learned from a monocular video.
• Generates arbitrary expressions by linearly blending neutral and blendshape Gaussians with expression coefficients, mirroring classical parametric face models.
• Real-time synthesis via Gaussian splatting, with claimed better high-frequency detail and rendering performance than prior methods.

Context

Bridges classical blendshape control (cf. Lewis and Anjyo, 2010, Direct Manipulation Blendshapes) with rigged Gaussian-splatting avatars (cf. GaussianAvatars, Qian et al., 2024).Builds on: Direct Manipulation Blendshapes · GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians

Correctness

Trained and demonstrated from monocular video with comparisons to state-of-the-art; quality depends on input-video coverage of expressions, and linear Gaussian blending is an approximation that may struggle with expressions far outside the learned basis.

Clarity

Accessible to readers familiar with blendshapes and 3DGS; a first pass conveys the mapping, a second pass is needed for the Gaussian-property blending details.

How to read it

Read for how classical blendshape semantics map onto Gaussians and the real-time blending; second pass on training and per-Gaussian parameters if you intend to build or animate such avatars.