Spaces:
Running on CPU Upgrade
Running on CPU Upgrade
| title: Sand3D Explorer | |
| colorFrom: gray | |
| colorTo: purple | |
| sdk: docker | |
| app_port: 7860 | |
| pinned: false | |
| short_description: 3D objects, each with the Python program that builds it | |
| # Sand3D Explorer | |
| Browse Sand3D, Sandflow's open dataset of 3D objects written as code: every object comes with | |
| the short [sand3d](https://github.com/Sandflow-ai/sand3d) program that builds it. Filter by | |
| dataset, family or joints, search by words or by look, read and copy a program, and open the | |
| object in 3D where a GLB was built for it. | |
| The programs come from Sandflow's iterative agentic reconstruction pipeline. Each was run again | |
| under sand3d, in a fresh process, and kept only when it rebuilt the stored object. | |
| ## How it runs | |
| A FastAPI process serves one static page and a JSON API, on the free CPU tier. At start it | |
| downloads the datasets listed in `config.json` (the `HF_TOKEN` secret reads them), keeps their | |
| metadata and programs in memory and their renders in one file on disk, then reads | |
| `Sandflow/sand3d-explorer-assets`: one GLB per object for a varied sample, and one image | |
| embedding per render. A text query is encoded on CPU by the same open image-text model the | |
| renders were embedded with. | |