Gaussian Deja-vu

01 — Overview

Overview

Personalized head avatars often require a lengthy per-person optimization process. Gaussian Deja-vu targets both generalization to a new identity and efficient personalization while preserving explicit control over expression and pose.

The method uses 3D Gaussian rendering to retain real-time performance and high-frequency appearance detail.

02 — Contributions

Key Contributions

• 01

  Combines generalizable initialization with efficient identity-specific personalization.

• 02

  Builds controllable head avatars around a real-time 3D Gaussian representation.

• 03

  Improves personalized avatar quality while reducing the time needed to adapt to a new subject.

03 — Method

Method

Gaussian Deja-vu learns reusable priors across identities and then adapts the Gaussian avatar representation to a target subject. Facial controls drive the personalized representation while the Gaussian renderer produces novel views in real time.

04 — Evaluation

Results

The WACV evaluation reports faster personalization and improved photorealistic avatar quality, together with controllable expression and pose rendering.

05 — Reference

Citation

@InProceedings{Yan_2025_WACV,
  author    = {Yan, Peizhi and Ward, Rabab and Tang, Qiang and Du, Shan},
  title     = {Gaussian Deja-vu: Creating Controllable 3D Gaussian Head-Avatars with Enhanced Generalization and Personalization Abilities},
  booktitle = {Proceedings of the Winter Conference on Applications of Computer Vision (WACV)},
  month     = {February},
  year      = {2025},
  pages     = {276--286}
}