nv-tlabs/XCube

What it solves

XCube addresses the challenge of generating high-resolution 3D voxel grids without the need for time-consuming test-time optimization. It enables the creation of detailed 3D objects and large-scale outdoor scenes (up to 100m x 100m) with high effective resolution (up to 1024³), allowing for millions of voxels to be generated in a feed-forward manner.

How it works

XCube uses a hierarchical voxel latent diffusion model. It generates 3D grids progressively from coarse to fine resolutions using a custom framework built on the efficient VDB data structure. This approach allows the model to handle sparse 3D voxel grids with arbitrary attributes.

Who it’s for

This project is designed for researchers and developers working in 3D generative AI, spatial intelligence, and high-resolution 3D scene synthesis.

Highlights

• High Resolution: Supports a finest effective resolution of up to 1024³.
• Feed-forward Generation: Generates 3D content without requiring test-time optimization.
• Versatile Tasks: Capable of unconditional generation, user-guided editing, scene completion from a single scan, and text-to-3D.
• Large Scale: Demonstrated effectiveness on large outdoor scenes with voxel sizes as small as 10cm.