Instance-Structured 3D Tokenization via Unposed Multi-View Images
June 27, 2026
This framework decomposes 3D scenes into object-centric token groups directly from unposed images. By pairing instance tokens with local geometry anchor tokens, it enables downstream reconstruction, segmentation, and manipulation through a two-level factorization of identity and appearance.
HOW THIS AFFECTS YOU
●
researcherThis provides a more structured alternative to unstructured Gaussian or point cloud outputs.
●
designerThis enables more intuitive object-level manipulation within 3D generated scenes.