1 paper · 2 filters
Xiang Chen, Alexander Binder
Current 3D tokenizers largely treat representation as spatial compression: compact codes reconstruct surface geometry, but leave component ownership and attachment validity implici…