The framework adapts the Pixal3D single-object prior with spatially grounded multi-view conditioning. DINOv3 features are projected into each object canonical voxel grid and fused across views. Generated objects are assembled using rigid transforms, allowing scene construction without cross-object mesh fusion or per-object post-processing optimization.
WorldSculpt supports editable reconstruction and conversion of generated Gaussian-splat worlds into mesh assets. The release includes code, models, and benchmark resources. Ground-truth meshes for UE-MeshyScene have redistribution restrictions, so benchmark evaluation may require contacting the authors rather than downloading every asset directly.
