Sand3D: 106,094 3D objects in one code substrate for language models
Blog post from Hugging Face
Sand3D is a collection of 106,094 3D objects represented as short Python programs using Sandflow’s open-source sand3d library, including 10,456 articulated objects drawn from Objaverse, CAD, Articraft, AutoMate, and NVIDIA SimReady datasets. Each entry includes executable source code, an exported OpenUSD asset, and a render, with programs designed around consistent named parts, parameterized dimensions, parent-relative placement, and explicit joints rather than the varied conventions of existing 3D tools. The project argues that this structured, recipe-like representation uses fewer tokens than Blender Python, helps language models plan objects before generating detail, supports editing and regeneration, and produces cleaner, part-aware geometry than mesh-based source datasets. Programs were generated through an iterative reconstruction pipeline and retained only when they met source-specific F-score thresholds, with articulated assets also required to reconstruct each part and joint accurately. The initial release focuses on hard-surface objects such as furniture, tools, machinery, and appliances, while future work aims to support organic forms through additional geometric primitives and deformation methods. The dataset is available under CC BY 4.0 subject to original source licenses, while the sand3d library is released under Apache 2.0.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| LLM | 2 | 3,355 | 742 | 191 | -49% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.