Generalization to dramatically novel views for Gaussian blendshape avatars

Improve the generalization capability of the 3D Gaussian blendshape head avatar representation, trained from monocular videos and rendered via 3D Gaussian splatting with FLAME-based control, to robustly handle dramatically novel viewpoints (e.g., extreme side views) when such views are absent from the training data, thereby mitigating side-view rendering artifacts observed across Gaussian- and NeRF-based head avatar methods.

Background

The paper introduces a 3D Gaussian blendshape representation for head avatars that enables real-time, high-fidelity animation by linearly blending Gaussian blendshapes according to expression coefficients and rendering via Gaussian splatting. Despite strong performance, the authors report failure cases when rendering extreme side views if such views are not present in the training set.

This limitation is shared with prior NeRF- and Gaussian-based head avatar approaches, highlighting a broader challenge of view generalization. Addressing this open problem is critical for robust deployment in applications requiring reliable appearance under large viewpoint changes, such as telepresence, VR/AR, and cross-identity reenactment.