maxwells-daemons / cppn

Playing around with compositional pattern-producing networks

CPPN

Playing around with generative art through compositional pattern-producing networks, neural networks that map (x, y) -> (r, g, b).

Initially written in January 2021 to make desktop backgrounds. Includes code and demos of symmetric CPPNs, using bitmaps as inputs, and animations, including looping.

Files

• cppn_random.ipynb: Random fully-connected networks, including enforcing symmetries.
• cppn_conv.ipynb: Random convolutional networks.
• cppn_paramcircle.ipynb: Using a controllable input channel to add structured patterns to the CPPN output.
• cppn_image.ipynb: Using an image bitmap as the input coordinate, effectively using the CPPN as a nonlinear colormap for an existing raster.
• cppn_time.ipynb: Making animations by running a CPPN with a time dimension. Animations can be looped by walking parameters around a high-dimensional circle.
• cppn_styletransfer.ipynb: Optimizing a CPPN to match the style of a target image. This one doesn't really work.