What is it about?

Our work explores how modern AI can bridge the gap between human language and 3D geometry. Open3DSearch allows users to locate 3D shapes by typing free-form text descriptions instead of relying on keywords or categories. It uses large vision-language models, originally trained for images and text, to reason about how a written description corresponds to visual views of 3D objects. This makes 3D search more flexible and accurate than traditional methods.

Featured Image

Why is it important?

As 3D content explodes across VR/AR, design, and digital entertainment, being able to search 3D shapes with free-form text is crucial but still largely unsolved. Open3DSearch is the first framework to enable zero-shot 3D shape retrieval in open domains—without needing any category-specific training data. This approach opens a new direction for building general-purpose 3D search engines, democratizing access to 3D assets and accelerating creation in design, manufacturing, and virtual worlds.

Perspectives

It’s inspiring to think that designers will soon search and reuse 3D shapes as easily as typing a sentence, making creativity faster, more intuitive, and truly language-driven.

Xiong Li
Zhejiang University of Technology

Read the Original

This page is a summary of: Open3DSearch: Zero-Shot Precise Retrieval of 3D Shapes Using Text Descriptions, October 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3746027.3755533.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page