What changed
Hugging Face has announced the release of the Sand3D dataset, which includes 106,094 3D objects, each represented as a short Python program. The dataset is built using the sand3d library, an open-source tool that allows language models to learn 3D structures through a unified code format. The objects are sourced from various repositories, including Objaverse, CAD parts, articulated mechanisms, and assemblies. Each object comes with its corresponding Python code, an OpenUSD file, and a render.
What this means in practice
The Sand3D dataset is designed to simplify the learning process for language models by providing a consistent structure for 3D objects. Each object is described as a recipe, with named parts, parameters, and placement details. This structured approach allows models to generate and modify 3D objects more efficiently. The dataset includes 10,456 articulated objects, which are particularly useful for applications requiring motion and interaction. The sand3d library is available on PyPI, and its documentation can be found at sand3d.sandflow.ai.
Limitations and availability
The current release of Sand3D focuses on hard-surface objects such as furniture, tools, and mechanical parts. Organic shapes, characters, and animals are not included and may require new verbs or methods in the library. The dataset is released under the CC BY 4.0 license, with each object retaining the license of its source. The sand3d library is available under the Apache 2.0 license. The dataset is accessible through the Hugging Face Hub, and the Explorer tool allows users to browse and interact with the objects and their code.
