mlx-spatial

MLX-native 3D and spatial inference tooling for Apple Silicon.

mlx-spatial is a practical runtime package for running modern 3D reconstruction and image-to-3D pipelines locally with MLX. The package is intentionally focused: keep weights outside the wheel, validate the assets you downloaded, then run clear command-line paths that produce inspectable outputs.

This is not a training framework, and it does not bundle model weights.

What Works Now

The package covers five model families:

Pipeline	Input	Output	Weight setup
SAM 3D Objects	image + object mask	Gaussian PLY, optional GLB	appautomaton MLX bundle
TRELLIS.2	object-centric RGB/RGBA image	shape OBJ or textured GLB	downloaded safetensors directly
HY-WorldMirror 2.0	scene image or image frames	camera, depth, normals, point-cloud PLY	downloaded safetensors directly
LiTo	object-centric RGB/RGBA image	3D Gaussian Splat PLY	appautomaton research MLX bundle
MapAnything	scene image views	scene .npz with depth, cameras, and world points	downloaded safetensors directly

Choose by job:

• Use SAM3D when you have an object image plus an exact mask and want object reconstruction with Gaussian PLY output.
• Use TRELLIS.2 when you have an object-centric image and want a shape OBJ or textured GLB.
• Use HY-WorldMirror when the input is a scene or frame set and you need camera, depth, normal, or point-cloud outputs.
• Use LiTo when you want Apple's research image-to-3DGS path and can work with Gaussian splat PLY output instead of a mesh.
• Use MapAnything when you have related scene views and want image-only depth, confidence, masks, camera parameters, and dense world points.

Honest status:

• SAM3D is the strongest object reconstruction path in this package. It uses the public appautomaton/sam-3d-objects-mlx bundle.
• TRELLIS.2 generation works, including textured GLB export. The export path is usable, but still an area we keep improving for texture and mesh quality.
• HY-WorldMirror works for scene reconstruction with camera,depth,normal,points. The optional Gaussian head is not part of the release-ready path yet.
• LiTo runs checkpoint-backed image-to-3DGS inference with the public appautomaton/lito-research-mlx bundle. Outputs are Gaussian splat PLY files, not meshes; use a 3DGS-aware viewer.
• MapAnything runs checkpoint-backed scene generation with the public facebook/map-anything weights. The supported artifact is a scene .npz tensor bundle, not a mesh or Gaussian splat export.

Install

For local development from this repo:

uv sync
uv run pytest -q

For package consumers:

uv add mlx-spatial
# or
pip install mlx-spatial

Requirements:

• Python 3.11+
• Apple Silicon recommended
• MLX installed through the package dependencies
• model weights downloaded separately under weights/

Command Line Tools

The package installs five CLIs:

uv run mlx-spatial-sam3d --help
uv run mlx-spatial-trellis2 --help
uv run mlx-spatial-hyworld2 --help
uv run mlx-spatial-lito --help
uv run mlx-spatial-mapanything --help

The repository also includes readable script wrappers under scripts/. These are the easiest starting point because they encode recommended settings.

Model Assets

Weights are intentionally not committed and not shipped in the wheel. Keep them under ignored local folders:

weights/sam-3d-objects-mlx/
weights/lito-research-mlx/
weights/trellis2/
weights/rmbg2/
weights/dinov3-vitl16-pretrain-lvd1689m/
weights/hy-world-2/
weights/map-anything/

SAM3D uses the converted appautomaton/sam-3d-objects-mlx runtime bundle:

uv run hf download appautomaton/sam-3d-objects-mlx \
  --local-dir weights/sam-3d-objects-mlx
uv run mlx-spatial-sam3d validate weights/sam-3d-objects-mlx

LiTo uses the converted appautomaton/lito-research-mlx research bundle:

uv run hf download appautomaton/lito-research-mlx \
  --local-dir weights/lito-research-mlx
uv run mlx-spatial-lito validate weights/lito-research-mlx

TRELLIS.2, HY-WorldMirror, and MapAnything do not need SAM3D-style conversion. They load the downloaded safetensors and JSON configs directly:

uv run mlx-spatial-trellis2 download-command --root weights/trellis2
uv run mlx-spatial-trellis2 rmbg-download-command --root weights/rmbg2
uv run mlx-spatial-trellis2 dinov3-download-command weights/dinov3-vitl16-pretrain-lvd1689m
uv run mlx-spatial-hyworld2 download-command weights/hy-world-2
uv run mlx-spatial-mapanything download-command weights/map-anything

Run the printed hf download ... commands, then validate:

uv run mlx-spatial-trellis2 validate --root weights/trellis2
uv run mlx-spatial-trellis2 rmbg-validate --root weights/rmbg2
uv run mlx-spatial-trellis2 dinov3-validate weights/dinov3-vitl16-pretrain-lvd1689m
uv run mlx-spatial-hyworld2 validate weights/hy-world-2
uv run mlx-spatial-mapanything validate weights/map-anything

Respect the licenses and access terms of the upstream model providers. The Python package only provides runtime code.

First Runs

SAM3D Object Reconstruction

Use an image and the exact object mask you want reconstructed:

python scripts/sam3d/reconstruct.py inputs/sam3d/living-room/image.png \
  --mask inputs/sam3d/living-room/mask-3.png \
  --output-dir outputs/sam3d/living-room-script

Expected output:

outputs/sam3d/living-room-script/
  gaussians.ply
  trace.json

Inspect the trace:

python scripts/sam3d/inspect_trace.py outputs/sam3d/living-room-script/trace.json

TRELLIS.2 Textured GLB

Use an object-centric image. RGBA images use their alpha channel directly; RGB images use RMBG to estimate the foreground:

python scripts/trellis2/generate_textured.py inputs/trellis2/cup-of-tea.jpg \
  --output-dir outputs/trellis2/cup-of-tea-script

Expected output:

outputs/trellis2/cup-of-tea-script/
  model.glb
  trace.json

The default settings are quality-oriented for Apple Silicon: 512 pipeline, model-config sampler steps, 1024 texture, 200k GLB face target, global xatlas unwrap, and kdtree texture baking. Low-step runs are useful for smoke tests, but they are not representative of output quality.

HY-WorldMirror Scene Reconstruction

Use a scene image or a directory of scene frames. This pipeline does not take an object mask:

python scripts/hyworld2/generate_scene.py inputs/sam3d/kidsroom/image.png \
  --output-dir outputs/hyworld2/kidsroom-scene-script

Expected output:

outputs/hyworld2/kidsroom-scene-script/
  camera_params.json
  depth/
  normal/
  points/points.ply
  trace.json

The script uses the verified release path: real Tencent safetensors, large memory profile, and camera,depth,normal,points heads. For frame directories, use --memory-profile balanced when the large profile hits the attention guard.

LiTo Image to 3D Gaussian Splat

Use an object-centric image with a useful alpha mask when possible:

python scripts/lito/generate.py inputs/lito/sample.png \
  --weights-root weights/lito-research-mlx \
  --output outputs/lito/sample.ply \
  --memory-profile balanced \
  --print-metrics

Expected output:

outputs/lito/sample.ply
outputs/lito/sample.safetensors

LiTo writes a Gaussian Splat PLY, not a mesh. Blender's native PLY importer can read the container, but it does not render the 3DGS fields correctly. Use a Gaussian-splat-aware viewer such as KIRI's Blender 3DGS add-on.

MapAnything Scene Bundle

Use a directory of related scene views. The Desk example is a two-image scene:

python scripts/mapanything/generate_scene.py inputs/map-anything/desk \
  --output-dir outputs/mapanything/desk-script

Expected output:

outputs/mapanything/desk-script/
  scene.npz
  trace.json

The script uses the upstream image-only inference settings: fixed_mapping preprocessing, stride 1, checkpoint-derived patch size, DINOv2 normalization, and mask/edge-mask postprocessing. scene.npz matches the original Torch scene layout semantically: images, depth, confidence, masks, intrinsics, camera poses, and world points. The MLX file uses clean top-level keys and also records extrinsics.

Repository Layout

src/mlx_spatial/     package code
scripts/             readable user and maintainer wrappers
docs/                deeper setup, release, and architecture notes
tests/               unit and parity-oriented coverage
weights/             ignored local model assets
inputs/              ignored local sample inputs
outputs/             ignored generated results
vendors/             ignored upstream checkouts

Documentation

• docs/README.md: documentation map and reader contract.
• scripts/README.md: recommended inference scripts and their defaults.
• docs/sam3d.md: SAM3D setup, inference, quality gates, PLY expectations, and coordinate notes.
• docs/trellis2.md: TRELLIS.2 asset layout, no-conversion note, scripts, and export caveats.
• docs/hyworld2.md: HY-WorldMirror asset layout, scene inputs, memory profiles, and outputs.
• docs/lito.md: LiTo setup, research-weight bundle, image-to-3DGS CLI, memory profiles, and PLY viewing notes.
• docs/mapanything.md: MapAnything asset layout, scene .npz schema, parity notes, and viewer/export boundary.
• docs/architecture.md: module map and pipeline boundaries.
• docs/development.md: tests, local asset rules, and contribution constraints.
• docs/model-publishing.md: model bundles and model-card rules.
• docs/release.md: release checklist.

Release Hygiene

Before publishing, build and inspect the artifacts:

uv run pytest -q
rm -rf dist
uv build
python scripts/packaging/check_release_artifacts.py \
  dist/mlx_spatial-*.tar.gz \
  dist/mlx_spatial-*-py3-none-any.whl
python scripts/packaging/check_release_artifacts.py --git-hygiene

The build must not include local weights, generated outputs, inputs, vendor checkouts, caches, or agent state.

Publishing is handled by the trusted-publishing workflow in .github/workflows/workflow.yaml. Do not publish from local shell credentials.