VivaHead: generation-augmented Gaussian avatars for single-image full-head reconstruction and animation
File(s)
Published version
Author(s)
Luo, Yiming
Yang, Molan
Ghosh, Abhijeet
Type
Conference Paper
Abstract
Animating a complete 3D head avatar from a single portrait image is highly challenging. A single frontal image provides no evidence for the side or back of the head and contains only one static expression, yet the recovered avatar must be geometrically complete, identity-consistent, and faithfully drivable under arbitrary expressions. While animatable avatar methods have advanced rapidly, most existing approaches still focus on frontal faces rather than the entire head. Conversely, existing full-head methods are largely non-animatable. The few methods that attempt to support both full-head reconstruction and animation are still limited, often rely on mesh-based representations, and may struggle to produce plausible back-head geometry or reliably preserve facial identity. As a result, methods that can jointly deliver high-quality full-head reconstruction, particularly in unseen back-head regions, and high-fidelity, vivid animation remain very rare.We present VivaHead, a generation-augmented Gaussian avatar framework for single-image full-head reconstruction and animation. Our method compensates for the severe information deficiency of a single portrait with two complementary types of synthesized observations. Specifically, we generate multi-view images under predefined viewpoints to provide additional coverage of the side and back head, and generate multi-expression images of the same identity via diffusion-based reenactment to recover richer expression-dependent facial details absent from the source image. Together, these synthesized observations provide more informative cues for both full-head completion and animation detail recovery. The source image, together with the generated multi-view and multi-expression images, is fed into a feed-forward Gaussian reconstruction branch that predicts an initial canonical animatable full-head avatar in a single forward pass. We further adopt an expression-neutral representation that suppresses source-expression leakage into the canonical avatar, reducing residual source-expression artifacts in the final animation and enabling more faithful driver-conditioned animation. Starting from this initial prediction, we apply a lightweight refinement stage that leverages the synthesized observations to improve geometric fidelity and appearance details.
Date Issued
2026-07-01
Date Acceptance
2026-06-10
Citation
Rendering 2026 Industry Track, 2026
ISBN
978-3-03868-320-9
ISSN
1727-3463
Publisher
The Eurographics Association
Journal / Book Title
Rendering 2026 Industry Track
Copyright Statement
© 2026 The Author(s). Proceedings published by Eurographics - The European Association for Computer Graphics. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.
License URL
Source
EGSR 2026
Publication Status
Published
Start Date
2026-07-01
Finish Date
2026-07-03
Coverage Spatial
Bordeaux, France