World from Motion accepted to NeurIPS 2026, with code and model weights released
World from Motion uses single-camera video and an initial reconstruction to generate plausible views of a dynamic 3D scene.
TLDR
Researchers have released code, model weights and DyCheck reproduction scripts for World from Motion, a Stanford-NVIDIA collaboration. The system folds generated views from new camera positions into a dynamic 3D reconstruction. Those views are plausible completions, not observations; unseen areas may be wrong or temporally inconsistent. The team's evaluation covered seven DyCheck iPhone scenes.
Combined views
4.7K
3 Sources, first seen ago