AnyAvatar: High-Fidelity Gaussian Head Avatars under Uncalibrated Camera Settings
Abstract
Most Gaussian head avatar methods rely on accurately calibrated multi-view cameras. Under uncalibrated settings, where camera poses are obtained from image-based prediction, inaccurate pose initialization poses a major challenge across two coupled stages. During FLAME fitting, pose errors can undermine cross-view head observability and destabilize the geometric prior, while residual errors during Gaussian Splatting (GS) optimization can introduce projection inconsistency, leading to training-view overfitting and degraded novel-view quality. To this end, we propose AnyAvatar, a unified framework for high-fidelity and controllable Gaussian head avatars under uncalibrated settings. First, we introduce Feasibility-Guided FLAME Initialization, which localizes a feasible FLAME position from coarse predicted poses through Constrained Ray Proximity Localization, restoring cross-view head observability and providing a reliable geometric prior. We then introduce Pose-Driven Gaussian Optimization, which jointly refines camera poses, Gaussian attributes, and FLAME parameters during GS training, while using a structured triplane-based appearance module to suppress pose-error compensation and improve cross-view consistency. Experiments on multiple uncalibrated datasets and self-captured data demonstrate state-of-the-art performance and robustness across diverse camera pose predictors.