ARToolKitNFT for python

This is a python binding for WebARKitLib library. It is based on the WebARKitLib library and provides a python interface to the library. Install from PyPI:

pip install artoolkitnft

Supported platforms

There is deliberately no source distribution: the build compiles sources from outside the package directory and needs the WebARKitLib git submodule, neither of which survives an sdist. So on a platform without a wheel, pip reports no matching distribution found rather than attempting a build that cannot succeed.

Intel macOS is not supported. GitHub retired its free Intel runner (macos-13) in December 2025; the replacement is a paid runner, and GitHub removes x86_64 macOS entirely in August 2027. Apple has already discontinued the architecture. If you need Intel macOS, build from a repository checkout following Local development below, or open an issue -- cross-compiling universal2 wheels is possible if there is demand.

Rehearsal builds are published to TestPyPI by running the Publish Python package workflow manually with the testpypi target. Note that any python/* tag publishes to PyPI, including a pre-release version such as 0.0.14rc1 -- routing is chosen by how the workflow is triggered, not by the version number. To install a rehearsal build:

pip install -i https://test.pypi.org/simple/ artoolkitnft

Local development (build and install from source)

To build the bindings locally and test them without publishing to TestPyPI:

Prerequisites

• Python 3.9+ with pip
• A C/C++ toolchain (MSVC Build Tools on Windows, build-essential on Linux)
• cmake available on PATH (used by setup.py to build the bundled zlib)
• The git submodules initialised:

  git submodule update --init
  

• On Windows: vcpkg with libjpeg-turbo and pthreads installed (see the build-windows job in .github/workflows/build-python.yml for the exact commands).
• On Linux: sudo apt-get install -y libjpeg9

Build and install

From the python-bindings/ directory:

pip install --upgrade pip setuptools wheel pybind11 numpy pillow pytest
python setup.py bdist_wheel
pip install --force-reinstall dist/artoolkitnft-*.whl

On Windows PowerShell the last step needs an explicit Get-ChildItem because PowerShell does not expand globs the way bash does:

pip install --force-reinstall (Get-ChildItem dist\artoolkitnft-*.whl | Select-Object -First 1).FullName

--force-reinstall is important — without it, pip sees the version has not changed and skips reinstalling, so your C++ changes never land.

Run the example

python example.py

Notes

• Do not use python -m build here: it creates an isolated build environment that does not inherit cmake (or vcpkg) from your shell. Use the legacy python setup.py bdist_wheel instead, which keeps your current PATH.
• Working inside a virtualenv (python -m venv .venv) is recommended so you do not pollute the system / Anaconda site-packages.

Publishing

Releases are built and published by the Publish Python package workflow, through PyPI trusted publishing (OIDC). There is no token, and nothing is uploaded by hand.

1. Bump the version in both pyproject.toml and setup.py — the workflow refuses to publish if the tag and the two files disagree.
2. Rehearse: run the workflow manually with the testpypi target and confirm the wheels build on all three platforms.
3. Release: tag python/<version> (for example python/0.0.13) and push it.

Do not publish by hand

This section previously documented a manual twine upload. It has been removed because it produced a broken artifact that is still on TestPyPI. The instruction was:

# DO NOT DO THIS
python setup.py sdist bdist_wheel --plat-name manylinux2014_x86_64
twine upload --repository-url https://test.pypi.org/legacy/ dist/*

--plat-name forcibly stamps a platform tag onto whatever was just built, without checking it. Run on Windows, or with a stale build tree, it labels a Windows .pyd as Linux. That is exactly what artoolkitnft-0.0.12-cp311-cp311-manylinux2014_x86_64.whl on TestPyPI contains: artoolkitnft_core.cp311-win_amd64.pyd, which installs on Linux and then fails to load.

It also uploads dist/*, which includes an sdist — and a source build cannot work here, because it needs sources from outside the package directory and the WebARKitLib submodule.

The workflow avoids all of this: it builds natively per platform, repairs wheels with auditwheel / delocate, rejects any wheel whose contents disagree with its tag (.github/scripts/check_wheel.py), installs and imports each one in a clean virtualenv, and publishes nothing unless every wheel in the matrix succeeded.