HeadsUp reconstructs high-quality 3D heads from large multi-View Captures

A new paper published in May 2026 introduces HeadsUp, a scalable feed-forward method for reconstructing detailed 3D Gaussian head models from large multi — camera captures. The work demonstrates end-to-end training and inference without per-subject fitting, aiming to make high-fidelity 3D head reconstruction practical for large capture pipelines and faster downstream production use.

HeadsUp uses an efficient encoder — decoder pipeline: multiple input views are compressed into a compact latent representation, which the decoder maps to a set of UV-parameterized 3D Gaussians anchored to a neutral head template. The UV parametrization explicitly decouples the number of 3D Gaussians from the number and resolution of input images, allowing the model to absorb many high-resolution views during training without inflating the Gaussian count.

The authors trained and evaluated HeadsUp on an internal multi — view dataset containing more than 10,000 distinct subjects — an order of magnitude larger than the multi — view human head datasets cited in the paper. Reported results claim state — of-the-art reconstruction quality and the ability to generalize to novel identities without any test-time optimization or per-subject fitting. Beyond per-instance metrics, the paper presents an extensive scaling analysis that probes behavior across identity count, view count, and model capacity. Those experiments expose practical quality — versus-compute trade — offs and offer guidance on where to allocate compute — whether to add more subjects, capture more views per subject, or increase model size-to reach target fidelity for downstream tasks.

Two concrete downstream demonstrations showcase the latent space: generating novel 3D head identities and animating reconstructed heads using expression blendshapes. Because HeadsUp runs feed-forward at test time and requires no per-scan optimization, these pipelines support faster inference and iteration than methods that rely on per-subject fitting. The operational implications for engineers and studios are practical: the UV-parameterized Gaussian representation combined with feed-forward inference makes it feasible to train on large, high-resolution capture farms and to deploy real-time or near-real-time reconstruction and animation workflows. The paper also situates HeadsUp relative to recent Gaussian — based and diffusion — driven 3D reconstruction research and includes methodological and authorship details.