GUSH3R: Single Forward Pass for Dynamic Human-Scene 3D Reconstruction

rsasaki0109 · x · 2026-08-04

Researchers have introduced GUSH3R, a feed-forward framework for online dynamic human-scene reconstruction. It tackles the challenging problem of reconstructing dynamic environments from monocular videos.

Previous feed-forward methods often predicted non-photorealistic representations (like point clouds and meshes) or failed to handle non-rigid objects, particularly dynamic humans. Utilizing 3D Gaussian Splatting (3DGS) primitives, GUSH3R jointly models dynamic humans and static scenes in a single forward pass, enabling photorealistic rendering of complex non-rigid motions.

Original post →

More from Multimodal

Multimodal channel →