GUSH3R Reconstructs Dynamic Humans and Static Scenes as 3D Gaussians from Monocular Video

Researchers at the University of Tokyo have introduced GUSH3R (Gaussian-Unified Scene Human 3D Reconstruction), a feed-forward framework that jointly reconstructs dynamic humans and static scenes from monocular video as 3D Gaussian Splatting primitives in a single forward pass. The system uses two dedicated decoder branches — a Scene Gaussian Decoder and a Human Gaussian Decoder — built on top of the foundation model Human3R, combining point cloud and human mesh priors to produce geometrically consistent, photorealistic novel view synthesis.

Unlike prior feed-forward methods that are typically limited to non-photorealistic representations such as point clouds or meshes, or that struggle with non-rigid objects like moving people, GUSH3R unifies human and scene reconstruction in a shared metric space. Experiments on monocular human-scene datasets show the approach achieves competitive novel view synthesis quality while significantly improving inference efficiency over optimization-based alternatives — a combination relevant to VFX, volumetric capture, and virtual production pipelines.