TensorFlow Graphics: one release from 2019 and a setup.py that rewrites its own source
TensorFlow Graphics: Differentiable Graphics Layers for TensorFlow
At a glance
- What is it?
- A library of differentiable graphics and geometry layers for TensorFlow whose only published release is v1.0 from May 2019, whose version number is stamped from the build date, whose input checks ship switched off, and whose requirements floor contradicts the compatibility line.
- Who is it for?
- TensorFlow Graphics is worth using for the layers it wraps, and it is not a place to look for a maintained dependency tree. Pin the version you test against, because the package you install from a nightly channel carries a date as its version and nothing has been published under the normal name since 2019.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 14 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- tensorflow/graphics2,782 ★ · +0
| Repository | 2026-10-04 | 2026-11-15 |
|---|---|---|
| tensorflow/graphics | 2.8k | 2.8k |
0 Stars in the last 42 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
One release, tagged 1.0.0, titled v1.0, published in 2019
The repository has a single published release. It is titled v1.0, its tag is 1.0.0 without the letter v, and it went out on 2019-05-09. The default branch has been pushed to since, with the last commit dated 2026-09-16, so the code has moved a long way from anything a package index can hand you. The tag and the title disagree on the same version, which is the only release where that mismatch is visible. The metadata also carries no homepage at all, so the repository page, the PyPI listing and the install document inside the tree are the three places to look. The compatibility paragraph names three TensorFlow channels rather than a version of this library, which is a sign of where the release process ended up.
The version number is the build date
setup.py does not read a version string out of the package. It computes one from the clock, formatting the current year, month and day, and passes that to setuptools as the version. The documented usage is a single line, python setup.py sdist bdist_wheel, with an optional --nightly flag. That flag does more than change a channel: it renames the distribution. Without it the package published to PyPI is tensorflow-graphics, and with it the name becomes tfg-nightly, so nightly and stable installs can sit in the same environment without one overwriting the other. The same date is used for both, which means two builds on the same day produce the same version number and the second one is the one that wins on the index.
Building rewrites a file inside the package source
Before it does anything else, the build script edits the tree. It locates tensorflow_graphics/__init__.py next to itself, reads the file into a list of lines, replaces the literal text __version__ = "HEAD" with the computed version, and writes the result back to the same path. The change is therefore not confined to a build directory: a plain sdist run leaves the source file modified on disk. The replacement is an exact string match, so a version line that has already been stamped, or written with different spacing, is passed over and the file keeps whatever it had. That is the quiet failure mode worth knowing about, since a stamped value left over from an earlier build and a deliberately pinned value look identical from the outside.
One path is resolved from the script, the other from the shell
The two file lookups in the build script disagree about where the project root is. The package init file is resolved from __file__, so it is found wherever the script lives, while the dependency list is opened as a bare relative path, requirements.txt, which the shell resolves against the current working directory instead. The comment at the top of the script states the assumption plainly: setup.py file lives at the root of the project. Run the command from anywhere else and the dependency read fails while the version rewrite has already happened. Dependencies are parsed by stripping each line and dropping the ones beginning with a hash, which is why the one comment line in the requirements file survives into the tree but not into the install list.
Fourteen dependencies across three TensorFlow packages
The declared requirements are wider than the name suggests. Alongside tensorflow itself, the list pulls in tensorflow-addons and tensorflow-datasets, so the version floor that matters is the intersection of three packages rather than one.
tensorflow >= 2.2.0
tensorflow-addons >= 0.10.0
tensorflow-datasets >= 2.0.0
absl-py >= 0.6.1
h5py >= 2.10.0
matplotlib >= 2.2.5
numpy >= 1.15.4
psutil >= 5.7.0
scipy >= 1.1.0
tqdm >= 4.45.0
OpenEXR >= 1.3.2
termcolor >= 1.1.0
trimesh >= 2.37.22
networkxnetworkx carries an inline comment marking it as required by trimesh rather than by this library directly. OpenEXR, h5py and matplotlib are the heavier entries, and the mesh library trimesh sits at a notably recent floor of 2.37.22 next to numpy at 1.15.4, so the pins are not maintained to one era. A second file, requirements.unix, holds the platform-specific remainder.
The compatibility line names a 2.0 preview, the dependency floor is 2.2.0
The compatibility paragraph promises support for the latest stable release of TensorFlow, for tf-nightly, and for tf-nightly-2.0-preview, adding that every function works under both graph and eager execution. The third of those channels belongs to the TensorFlow 2.0 era, while the requirements file asks for tensorflow at 2.2.0 or newer, so the sentence is describing a channel older than the library's own floor. Neither the version of this package nor a TensorFlow release number appears together anywhere else. The graph and eager claim is the more useful half, since it is the reason the layers can sit inside either style of model without a wrapper.
The input checks are off by default because they cost cycles
The debugging section is the most honest paragraph in the file, because it explains a wrong-answer mode rather than a crash. The library leans on L2 normalized tensors and on the inputs to specific functions sitting in pre-defined ranges. Verifying both takes cycles, so the verification is not activated by default, and the page recommends switching it on for a couple of epochs of training to confirm that everything behaves. Left off, a tensor that misses the expected range produces no error and no warning. The instructions for turning the checks on live in a separate debug_mode document in the tree, so the switch is not in the API surface.
Three dependency mechanisms and tutorials that disclaim themselves
The root carries three ways of resolving dependencies at once: a Bazel setup with WORKSPACE and a .bazelrc, a setuptools setup with setup.py and MANIFEST.in, and git submodules with .gitmodules and a submodules/ directory. Linting and test configuration sit alongside them in .flake8, .pylintrc and pytest.ini. The badge row repeats the PyPI link three times over, next to a coverage service link, an actions link, the Apache license mark and a download counter. The tutorial section runs against that grain: the Colab notebooks are ordered by difficulty and cover pose estimation, intrinsics optimization, interpolation, reflectance, non-rigid deformation, spherical harmonics and mesh convolutions, yet the page states they are not part of the API, can change at any time without warning, and should not be depended on in code.
Editorial conclusion
TensorFlow Graphics is worth using for the layers it wraps, and it is not a place to look for a maintained dependency tree. Pin the version you test against, because the package you install from a nightly channel carries a date as its version and nothing has been published under the normal name since 2019. Turn the debug checks on for the first epochs of any new setup, since the library assumes L2 normalized tensors and bounded input ranges without enforcing either by default. Read install.md rather than guessing at the install line, and treat the Colab notebooks as documentation that can change under you rather than as an API.
Frequently asked questions
How do I install TensorFlow Graphics?
The README gives no install command of its own and sends you to a separate install document in the repository, at tensorflow_graphics/g3doc/install.md. The build script's own usage line is python setup.py sdist bdist_wheel, with an optional --nightly flag that publishes the package as tfg-nightly instead of tensorflow-graphics.
What does TensorFlow Graphics provide?
Differentiable graphics and geometry layers, named as cameras, reflectance models, spatial transformations and mesh convolutions, plus 3D viewer functionality such as 3D TensorBoard. The layers are meant to be inserted into neural network architectures.
Which TensorFlow versions does TensorFlow Graphics support?
The compatibility section names the latest stable release of TensorFlow, tf-nightly, and tf-nightly-2.0-preview, and says all functions work under both graph and eager execution. The requirements file separately asks for tensorflow at 2.2.0 or newer.
Why does TensorFlow Graphics give wrong results without any error?
The library relies on L2 normalized tensors and on inputs to specific functions being in pre-defined ranges, and checking all of that takes cycles, so the checks are not activated by default. The page recommends enabling them for a couple of training epochs, with instructions in a separate debug_mode document.
Can I depend on the TensorFlow Graphics Colab tutorials?
No. The page states the tutorials are not considered part of the API, can change at any time without warning, and that writing code with a dependency on them is not advised.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/tensorflow-graphics)Star history chart for your README
[](https://hysenlabs.com/projects/tensorflow-graphics)Was this useful?