arxivcs.CVcs.AIcs.LG2026-07-07
ELSA3D: Elastic Semantic Anchoring for Unified 3D Understanding and Generation
Tianjiao Yu, Xinzhuo Li, Yifan Shen, Onkar Susladkar, Yuanzhe Liu, Xiaona Zhou, et al.
Unified 3D foundation models aspire to generate 3D assets and reason about them in language within a single backbone, but their text-3D interaction remains largely implicit. Existing methods concatenate text and 3D tokens into a flat sequence and rely on self-attention, collapsin…