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

FastGHA is a feed-forward method that builds animatable 3D Gaussian head avatars from only a few input images. It predicts Gaussian deformations from expression codes to support real-time animation.

FastGHA is a study on constructing animatable 3D head avatars from a small number of portrait images captured from different viewpoints and with different expressions. Existing methods often use large-scale multi-view capture equipment or require a separate optimization process for each person. This work proposes a feed-forward architecture that infers an avatar directly from input images so it can also be applied to previously unseen subjects.

The key idea is to learn a 3D Gaussian representation for every image pixel and combine information from multiple viewpoints in a transformer-based encoder. The researchers use image features from DINOv3 and the Stable Diffusion VAE together, while using the point maps of a large pretrained reconstruction model as geometric supervision to improve the smoothness of the head shape.

For animation, the method first reconstructs a canonical Gaussian head in a neutral state, after which a lightweight dynamic MLP predicts the 3D deformation of each Gaussian according to an expression code. This separates the expressions present in the input photos from the avatar’s movements, with the joint goal of reproducing fine-grained appearance and enabling real-time facial-expression animation. FastGHA was presented at ICLR 2026.

Maru