Screenshot of the GANDelve application

One of the most interesting models in the field of generative AI are GANs or Generative Adversarial Networks. This is a complex ensemble of 2 models that fight with each other, each improving in the process and creating new images, text or sound.

The mechanisms for GANs are complex and often not intuitive. Training GANs requires powerful machines that normal users do not have access to. In addition, they require coding and installation procedures that can be daunting.

To overcome these barriers, we developed GANDelve, an experimental tool for training Generative Adversarial Networks from scratch. The tool runs locally on users' computers, but uses Google's Cloud AI offerings to perform the actual training on fast GPUs hosted in Google's data centers.

GANDelve was introduced during LAbO 2019, ChampdAction's summer school that took place in late August 2019. We also had premaster and master students use the tool to train their own machine learning models.

A paper on GANDelve was published at the NeurIPS 2019 Creativity Workshop.


The following projects were made using GANDelve: