Generating 1024³ voxels without the test-time optimization slog

What it does

XCube is a generative model that outputs sparse 3D voxel grids with arbitrary attributes at resolutions up to 1024³. It handles everything from ShapeNet objects to 100 m × 100 m outdoor scenes with 10 cm voxels, and can perform unconditional generation, text-to-3D, user-guided editing, and scene completion from a single scan. The model runs feed-forward, meaning it skips the lengthy per-sample optimization loops that plague many neural 3D reconstruction pipelines.

The interesting bit

Instead of dense tensors, XCube builds on the VDB sparse data structure and fVDB deep-learning framework, using a hierarchical latent diffusion model that generates geometry coarse-to-fine. That sparse hierarchy is what lets it scale to millions of voxels without choking on empty space.

Key highlights

• Generates millions of voxels at up to 1024³ effective resolution in a single forward pass
• Handles large-scale outdoor scenes (100 m × 100 m at 10 cm voxel size) in addition to closed objects
• Supports unconditional generation, text-to-3D, scene completion from a single scan, and user-guided editing
• Built on sparse VDB hierarchies via the fVDB framework; requires an Ampere-or-newer GPU

Caveats

• Linux only; other platforms are not yet supported
• Waymo training data and single-scan conditioned inference are still marked “coming soon”
• The public release omits the paper’s refinement network and shifts mesh extraction to post-processing, which may produce slight variations from the published results

Verdict

Worth a look if you have a Linux workstation and a recent NVIDIA GPU, and need to generate 100 m × 100 m outdoor scenes or high-res objects without babysitting test-time optimization. Avoid if you require Windows, macOS, or a permissive open-source license.