Announcement
VoxLution generates 3D indoor scenes by refining coarse voxel layouts
The research team says its transformer refines scenes from 25 cm to 1.5625 cm voxels, capturing fine geometric detail.
TLDR
VoxLution first builds a low-resolution spatial layout, then refines scene geometry in local chunks. Its creators say it was trained on real-world scans, which are often noisy and partial. They report 70.5% lower FPD and 73.8% lower FID on the ScanNet++ benchmark than existing baselines.
Combined views
7.2K
3 Sources, first seen ago
