Daoyi Gao, Xinyi Zhang, Alexey Bokhovkin, Lei Li and Angela Dai (Technical University of Munich, University of Virginia) · 3D scene generation

VoxLution: Evolving Voxels for High-Fidelity 3D Scene Generation

VoxLution is a coarse-to-fine transformer method for generating 3D indoor scenes, from researchers at the Technical University of Munich and the University of Virginia. It is a sparse autoregressive cascade that refines a coarse scene layout from 25 cm down to 1.5625 cm voxels, aiming to avoid the overly smooth geometry of existing generative models. 3D super-resolution is treated as a sequence modelling problem over multi-resolution sparse voxel grids, using a training-free, resolution-invariant tokenisation that encodes the eight child octants of a voxel as one of 256 codes. Scenes are refined chunk by chunk, with neighbouring chunk boundaries included in the sequence for coherence, and a generative sparse VAE converts the final occupancy into a mesh. 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++, against existing baselines. It can also be conditioned on layouts of labelled bounding boxes. The paper is on arXiv and the code is listed as coming soon.