VoxLution

VoxLution is a research method for high-resolution 3D indoor scene generation. It uses a coarse-to-fine transformer approach: a coarse scene transformer first generates a low-resolution spatial layout as a token sequence, conditioned on scene type. Cascade transformers then refine the scene chunk by chunk, each stage doubling the resolution, down to 1.5625 cm voxels.

The method frames 3D super-resolution as sequence modeling over multi-resolution sparse voxel grids. It uses a training-free, resolution-invariant tokenization that encodes the eight octants of a voxel into a single decimal number. Adjacent chunk boundary information is included in the transformer sequence to keep local chunks coherent. Layout-conditioned generation is also shown.

The authors report 59.6% lower FPD and 51.1% lower FID on 3D-FRONT, and 70.5% lower FPD and 73.8% lower FID on ScanNet++, compared with existing baselines. Target uses include VR, spatial computing and robotics simulation. As of the provided material, only a project page is available. Code is described as coming soon, and no weights or hosted demo are listed. The work is by Gao, Zhang, Bokhovkin, Li and Dai, funded by an ERC Starting Grant and using EuroHPC supercomputers.