XCube

Why we included this project

XCube is a solid starting point if you want to generate high-resolution 3D voxel geometry without the slow per-sample optimization most diffusion models require. It builds on a hierarchical latent diffusion model over sparse voxel hierarchies, so a single feed-forward pass can emit millions of voxels at effective resolutions up to 1024³, and it scales to outdoor scenes of 100m × 100m with 10cm voxels. Beyond unconditional sampling, the released code covers text-to-3D on Objaverse, user-guided editing, and scene completion from a single scan, which makes it handy for prototyping asset pipelines. The repo ships pretrained checkpoints, inference scripts, and training configs for ShapeNet, Waymo, and Objaverse, so you can reproduce the paper or adapt the method to your own data. One caveat: it targets Linux with NVIDIA GPUs, and the code sits under NVIDIA's non-commercial source code license, so check that before building a commercial product on it.