FastGHA: Generalized Few-Shot 3D Gaussian Head Avatars with Real-Time Animation

TL;DR

FastGHA tackles scalable, high-fidelity 3D head avatar reconstruction from few-shot inputs by offering a feed-forward pipeline that yields animatable Gaussian head avatars. It reconstructs a canonical per-pixel Gaussian head using a multi-view transformer that fuses DINOv3 and SD-Turbo VAE features, augmented with per-Gaussian attributes, and then deforms Gaussians in real time via an MLP conditioned on FLAME expression codes. A VGGT-based geometry prior regularizes training to improve 3D head smoothness. Across Ava-256 and Nersemble, FastGHA achieves superior reconstruction quality and animation speed, with inference times under one second and real-time rendering, demonstrating strong generalization to unseen subjects and expressions.

Abstract

Despite recent progress in 3D Gaussian-based head avatar modeling, efficiently generating high fidelity avatars remains a challenge. Current methods typically rely on extensive multi-view capture setups or monocular videos with per-identity optimization during inference, limiting their scalability and ease of use on unseen subjects. To overcome these efficiency drawbacks, we propose \OURS, a feed-forward method to generate high-quality Gaussian head avatars from only a few input images while supporting real-time animation. Our approach directly learns a per-pixel Gaussian representation from the input images, and aggregates multi-view information using a transformer-based encoder that fuses image features from both DINOv3 and Stable Diffusion VAE. For real-time animation, we extend the explicit Gaussian representations with per-Gaussian features and introduce a lightweight MLP-based dynamic network to predict 3D Gaussian deformations from expression codes. Furthermore, to enhance geometric smoothness of the 3D head, we employ point maps from a pre-trained large reconstruction model as geometry supervision. Experiments show that our approach significantly outperforms existing methods in both rendering quality and inference efficiency, while supporting real-time dynamic avatar animation.

Figures (14)

• Figure 1: We present FastGHA, a feed-forward method that generates Gaussian head avatars from few-shot input images and animates them in real time.
• Figure 2: Overview of our method. Given a few input images with arbitrary views and expressions, we first extract multi-view features with pre-trained models and then train a multi-view transformer network that projects these features into 3D to reconstruct a canonical Gaussian head avatar. To enable real-time animation, we introduce a lightweight MLP that deforms the Gaussians according to the expression code.
• Figure 3: Qualitative reconstruction comparison on Ava-256 held-out subjects.
• Figure 4: Qualitative reconstruction comparison on Nersemble held-out subjects.
• Figure 5: Analysis on the different number of input views. Two inputs means the first two in the top row, three inputs means the first three, four inputs adds the first image in the bottom row, and so on.