Neural Art

Neural Art is a command line tool designed for fast and easy generation of abstract art using Compositional Pattern Producing Networks. The tool can be used to create either a single, or batch of random images using the underlying network.

The Numpy CPPN used to generate the image is directly based on the blog post by Nenad Markuš. I highly recommend reading the blog post as it gives a great introduction to CPPNs and the underlying math achieved with numpy. The post also goes on to create a very cool audio visualizer using CPPNs and Fourier Transforms.

Sample high resolution examples can be found in the /images/ folder.

 

Installation

This project is built using Python 3.9 and Pipenv. If you are unfamiliar with Pipenv, I recommend reading the official Pipenv documentation.

After cloning the project, navigate into the root directory and run:

pipenv shell
pipenv sync

This will create a new Python 3.9 virtual environment and install all neccesary dependencies from Pipfile.lock. If you have any issues while installing please refer to the Common Installation Issues at the bottom of this README.

 

Usage

The tool can be used to generate either a single image, or a batch of images. Both commands must be run from the root directory of the project.

Single Image Creation

The single command is designed for single image generation with tighter control over the underlying CPPN network. All flags have default values, more info can be found with the help flag.

python -m src.main single [-h] [-v] [-s SEED] [-l LAYERS] [-w WIDTH] [-r RESOLUTION RESOLUTION] [-a ACTIVATION]

Example

Generate a completely random image with a default resolution of 1920x1080.

python -m src.main single

Generate an image with a resolution of 5120x3600, numpy seed of 444, and a CPPN with 10 layers, width of 8 perceptrons, and a tanh activation function.

python -m src.main single -v -r 5120 3600 -s 444 -l 10 -w 8 -a tanh

 

Batch Image Creation

The batch command generates multiple images with random CPPN settings. The user must supply the number_of_images to be generated.

python -m src.main batch [-h] [-v] [-r RESOLUTION RESOLUTION] [-a ACTIVATION] number_of_images

Example

Generate 10 random images using the sech activation function

python -m src.main batch -a sech 10

 

Tips and Tricks

This tool is designed for experimentation! Therefore I encourage you to mess around with the different flags and the effects they have on the generated image.

A couple things:

• I recommend setting the resolution quite high. The image generation will be quite a lot slower, but the outputted images are generally much nicer (YMMV).
• Tanh is by far the best activation function for producing interesting art. I plan on adding functionality to use multiple activation functions, however I haven't figured out a nice way to achieve this with the current CLI.
• Have fun!

 

Tests

Running tests:

python -m unittest tests.[test_file_name]

 

Common Installation Issues

If unable to build numpy wheel for opencv-python package, you need to set the environment variable SYSTEM_VERSION_COMPAT=1. pypa/pipenv#4576.

 

References

Original Numpy network used here: Using CPPNs to generate abstract visualizations from audio data

Further CPPN Reading and Projects:

• Compositional Pattern Producing Networks: A Novel Abstraction of Development
• Understanding Compositional Pattern Producing Networks (Part One)
• Real time abstract art generation using a neural net
• Generating Abstract Patterns with Tensorflow