Our method reconstructs high-quality 3D faces from single images in just 10ms, regardless of input viewpoint.
In Submission
We present FastAvatar, a feed-forward framework for pose-invariant 3D face reconstruction from single images using 3D Gaussian Splatting. Our method directly predicts 3D Gaussian parameters in a single forward pass, eliminating the need for test-time optimization while handling arbitrary viewpoints.
We construct a canonical face model by averaging individual 3DGS models...
Our decoder predicts identity-specific offsets from the canonical model...
The encoder maps images from any viewpoint to consistent latent codes...
Single input view used to reconstruct the avatar.
In Submission
@article{liang2025fastavatar,
title={FastAvatar: Instant 3D Gaussian Splatting for Faces from Single Unconstrained Poses},
author={Liang, Hao and Ge, Zhixuan and Tiwari, Ashish and Majee, Soumendu and Godaliyadda, GM and Veeraraghavan, Ashok and Balakrishnan, Guha},
journal={arXiv preprint arXiv:2508.18389},
year={2025}
}