As part of our practice, we’re interested in pushing the boundaries of technology and exploring its applications for creative work. This study in particular explores the visual connections between noise maps and StyleGAN outputs with an attempt to use one as an output for the other.
Games with procedurally generated content, including Minecraft, Terraria, and No Man’s Sky, generally use noise maps for their random terrain generation. “Perlin noise” is a type of map that has the capability to generate realistic 2D and 3D terrain through its pseudo-random nature.
These are one dimensional (above) and two dimensional (below) visualizations of Perlin noise using p5’s built-in noise function. By graphing these values with lines and color, it becomes evident how they generate a smoothness in their randomness
The appearance of these noise maps were reminiscent of the early outputs you get from training a style GAN. We thought of the possibility of creating our own terrain with a noise map from a GAN model. We collected 100 different images of thunder as my dataset, curious to see if it could replicate the fine details of how the bolts fork in different directions.
This styleGAN generates a png of the training progress every 10 ticks which takes around 3 to 5 hours to render. It slowly goes from random colors and gradients into realistic lightning bolts. But at the end of this process, we realized that this actually looks nothing like Perlin noise.
Still, we wanted to know if it was possible to tie these images back into a noise map, so we attempted to recreate the lightning patterns with a Perlin and Worley noise