Start with the page that matches the job:
| Need | Page |
|---|---|
| Run SAM 3D Objects from an image and mask | sam3d.md |
| Run TRELLIS.2 object image-to-3D | trellis2.md |
| Run HY-WorldMirror scene reconstruction | hyworld2.md |
| Run Apple LiTo image-to-3DGS research inference | lito.md |
| Run MapAnything multi-view scene generation | mapanything.md |
| Inspect Pixal3D MLX implementation status | pixal3d.md |
| Start from recommended runnable scripts | scripts/README.md |
| Understand module boundaries | architecture.md |
| Contribute or verify changes locally | development.md |
| Publish converted model bundles | model-publishing.md |
| Prepare a PyPI release | release.md |
The docs are written for both humans and coding agents:
- Commands should be copyable from a clean checkout after
uv sync, or from an installed package when explicitly stated. - Model weights, inputs, and outputs stay under ignored local directories:
weights/,inputs/, andoutputs/. - Runtime pages should say what the pipeline is for, what assets it needs, how to run it, what it writes, and what common blockers mean.
- Avoid one-off run logs in stable docs. Put dated evidence in release notes, model cards, or local project notes.
- Keep upstream license and access terms linked, not pasted.
mlx-spatial ships runtime code and CLIs. It does not ship model weights.
Installed CLIs:
mlx-spatial-sam3d --help
mlx-spatial-trellis2 --help
mlx-spatial-hyworld2 --help
mlx-spatial-lito --help
mlx-spatial-mapanything --help
mlx-spatial-pixal3d --helpFrom a cloned repo, use uv run <command> for the same CLIs.