Large-Scale High-Quality 3D Gaussian Head Reconstruction from Multi-View Captures

HeadsUp is a feed-forward method that reconstructs high-quality 3D Gaussian heads in one shot from large-scale multi-camera input.

An efficient encoder-decoder compresses multiple views into a single latent representation, which is then decoded into UV-parameterized 3D Gaussians. These Gaussians are anchored to a neutral head template, creating a structure that is less tied to the number of input views and resolution.

The key effects are as follows.

• Using an internal dataset of more than 10,000 people, it secured a scale more than an order of magnitude larger than existing multi-view human head datasets.
• It generalizes to novel individuals without additional test-time optimization, achieving state-of-the-art reconstruction quality.
• By varying the number of identities, number of views, and model capacity, the scaling characteristics were analyzed to map out the quality-compute tradeoff.

The latent space was also used for generating novel 3D identities and for animation based on expression blendshapes.