WorldSculpt: Generating Compositional Worlds from Grounded Videos

Reconstructing heavily occluded, cluttered scenes typically yields a monolithic mesh with missing geometry in hidden areas. WorldSculpt circumvents this limitation by equipping a single-object generative prior (Pixal3D) with a multi-view conditioning pathway to ground generation in posed video observations. Trained entirely on isolated objects in canonical space without scene-level supervision, the method decomposes complex environments into hundreds of standalone object meshes and successfully converts raw 3D Gaussian Splatting captures into modular, interactive assets.