// the take
A small reference implementation of Compositional Pattern Producing Networks (CPPNs) in TensorFlow, used to generate abstract art and animated morphs by feeding coordinate grids through a randomly-weighted network. It's aimed at people who want to see how CPPNs work or tinker with generative art, not anyone building a production generative model.
The code is small enough to read end to end in one sitting (model.py, ops.py, sampler.py) which makes it a genuinely useful teaching example of how CPPNs map (x, y, r) inputs through alternating activations to produce images. It comes from hardmaru (David Ha), who wrote the original blog post explaining the technique, so the implementation matches the theory directly instead of being a guess at someone else's paper. The included GIFs and sample outputs make it easy to verify your own run produced something sane.
It targets TensorFlow 0.6/0.7, an API that has nothing in common with any TF1 or TF2 release since — you'd be rewriting the graph construction from scratch to run this today, not just patching import paths. There's no requirements file or version pin, so there's no way to know what else it depends on without reading the source. It hasn't been touched since 2018, there's no license file for the core code beyond a one-line mention in the README, and there are no tests, so any port is on you to verify correctness against the original rather than trusting CI.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →