Combining AI Tools to Create New Beetles with GANs #ArtificialIntelligence #MachineLearning #Colab #StyleGAN #RunwayML @cunicode
Earlier this year @cunicode announced the release of a model titled “Confusing Coleopterists” which generates new beetles! Well, images of beetles. This project started with a very thorough zoological guide to beetles titled, “Biologia Centrali-Americana :zoology, botany and archaeology“. The images in the book are beautiful, detailed and striking (see example excerpt above). @cunicode used the data in this book to create a dataset that was used to generate new AI-beetles. After experimenting with style transfer and deep dream @cunicode decided to work with GANs.
Previously I ran some test with DeepDream and StyleTransfer, but after discovering the material published at Machine Learning for Artists / @ml4a_ , I decided to experiment with the Generative Adversarial Network (GAN) approach.
The first iteration followed the methods in this lecture on utilizing DCGAN but generated some pretty fuzzy beetles. After that, @cunicode focused on StyleGAN. The first iteration utilized 128px images of beetles and trained the algorithm with PaperSpace. Three days, 1 GPU, and €125 later and some new bugs were created but they were a bit low res. The next step was to generate high-resolution beetles by loading the 1024px dataset of creepy crawlies onto RunwayML and train a model. The trained model was exported to Colab and used to generate never before seen beetles.
If you would like to try out this “buggy” model (we’re talking literal bugs, not digital ones) download RunwayML. If you want to tryout StyleGAN checkout this colab. If you would like to learn more about @cunicode’s methodology check out this post. Last but not least, if you want to get lost in images of AI-generated beetle’s take a look at cunicode’s Instagram or thisbeetledoesnotexist.com.