The authors introduce ELSA3D, a unified 3D foundation model that addresses implicit text-3D interaction by structuring language and geometric reasoning along matched abstraction scales. The model utilizes a scale-aware octree tokenizer and introduces Anchor Tokens, which are sparse cross-modal units that route semantic cues to relevant 3D scales.
- Anchor Tokens select semantic cues, retrieve scale-specific geometric evidence, and write the fused signal back into the unified representation.
- A lightweight per-block router makes computation and reasoning elastic by determining which text tokens instantiate anchors at specific geometric scales.
- ELSA3D achieves state-of-the-art performance in image-to-3D generation, text-to-3D generation, and 3D captioning.
- The approach halves FLOPs and inference latency compared to the non-elastic version of the same model.
This method allows cross-modal capacity to concentrate where alignment is most needed, improving efficiency without sacrificing performance.