Inspired by
Bart Hassink
Use image collections to train your own GAN
In 2022 Bart Hassink made a website that could generate dickpics using StyleGAN on RunwayML. A GAN, or Generative Adversarial Network, is a type of machine learning model invented in 2014 by computer scientist Ian Goodfellow. It’s a way of teaching a computer to generate new data that looks like other data (for example: images, sounds, or text). StyleGAN had pre-trained models on image-categories like landscapes or faces, but not this type of image. So Bart had to train his GAN from scratch. To do so he scraped the web for hundreds of publicly available dickpics as training material. Besides this generator, Hassink also trained an image recognition algorithm that could detect non-human phallic objects, for example in architecture or fauna.
Exercise
- Find an online tool that allows you to train a GAN of your own, for example Playform.io’s Freeform via https://playform.io/ (keep in mind that this one does cost money to use).
- Find or create one or multiple collections of similar images. For instance still life paintings from a certain painter or era from the online collection of the Rijksmuseum, mirror selfies taken at a certain place in your home from your personal photo archive, or royalty-free stock photos of produce. Make several versions of the collection with more and less images. For instance 30, 100 and 300 images.
- Train a GAN and let it create a series of new images with each version of the collection. If possible experiment with pre-trained models that can already recognize landscapes, for instance.
- Review what the differences are between the different outcomes in relation to the amount of images that were input or if a pre-trained model was used.