ELSA3D: Elastic Semantic Anchoring for Unified 3D Understanding and Generation

Forum topic · 小凯 · 2026-07-09

Summary

ELSA3D (arXiv:2507.06842) is a unified 3D foundation model that improves text-3D interaction through elastic semantic anchoring. Existing methods concatenate text and 3D tokens into a flat sequence and rely on self-attention, mixing coarse structural cues with fine geometric details. ELSA3D instead represents geometry with a scale-aware octree tokenizer and introduces Anchor Tokens—sparse cross-modal units that select semantic cues, route them to the most relevant 3D scale, retrieve scale-specific geometric evidence, and write fused signals back into the unified representation. A lightweight per-patch router decides which text tokens instantiate anchoring at which geometric scales, concentrating cross-modal computation where alignment matters most. The model achieves state-of-the-art results in image-to-3D generation, text-to-3D generation, and 3D captioning, outperforming the strongest unified baselines while cutting FLOPs and inference latency roughly in half compared to the non-elastic variant of the same model. The paper is by Tianjiao Yu, Xinzhuo Li, and Yifan Shen, posted July 9, 2025, in the computer vision category.

Paper Overview

Field: Computer Vision (CV)
Authors: Tianjiao Yu, Xinzhuo Li, Yifan Shen
Published: 2025-07-09
arXiv: 2507.06842

Original Abstract

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, collapsing coarse structural cues and fine geometric details into one undifferentiated representation.

We introduce ELSA3D, a unified 3D model that addresses this with elastic semantic anchoring, structuring language and geometric reasoning jointly along matched abstraction scales.

Key Ideas

Scale-aware octree tokenizer: represents geometry at multiple abstraction scales rather than as a flat token sequence.
Anchor Tokens: sparse cross-modal units that select semantic cues, route them to the most relevant 3D scale, retrieve scale-specific geometric evidence, and write the fused signals back into the unified representation—keeping interaction sparse and precise.
Elastic routing: a lightweight per-patch router decides which text tokens instantiate anchoring at which geometric scales, concentrating cross-modal capability where alignment is most needed.

Results

State-of-the-art performance in image-to-3D generation, text-to-3D generation, and 3D captioning.
Outperforms the strongest unified baselines.
Compared with the non-elastic version of the same model, FLOPs and inference latency are reduced by roughly half.