Given a subject mask in one view and a relational query such as "resting on", RelationVGGT segments the target object in every other view without being told its category. It runs in a single forward pass on unposed images, requiring neither per-scene optimization nor known camera poses.

3DGS segmentation renders accurate masks but selects an incomplete set of object Gaussians. VODA is a plug-and-play objective that directly supervises every Gaussian in an object's depth cluster, and Removal-3D benchmarks whether removing them actually changes the scene in 3D.

Generating diverse motions that also respect scene constraints is challenging due to the scarcity of paired datasets. SceneAdapt bridges separate scene–motion and text–motion datasets via motion inbetweening to inject scene awareness into text-to-motion models.

Can we freely explore reconstructed scenes from casually captured videos whose viewpoints are not sufficient? ExploreGS introduces a novel pipeline that samples virtual viewpoints to cover missing content and leverages diffusion priors, enabling explorable 3D scenes.

We propose a new framework for creating and easily manipulating 3D models of arbitrary objects using casually captured videos. Our core ingredient is a novel hierarchy deformation model, which captures motions of objects with a tree-structured bones.