Announcement
ELSA3D combines 3D asset generation and language descriptions in one model
The announcement says sparse Anchor Tokens route words to geometric scales, while an elastic router focuses cross-modal capacity.
TLDR
A post introducing ELSA3D says the model, led by Tianjiao Yu and accepted at NeurIPS 2026, generates 3D assets and describes them in language using one backbone. It says Anchor Tokens route words to geometric scales, while an elastic router directs capacity toward semantic–geometric alignment. The post claims state-of-the-art results in image-to-3D, text-to-3D and 3D captioning at roughly half the FLOPs and latency.
Combined views
611
2 Sources, first seen ago
