How artificial intelligence learned to design wheels
The name “FelGAN” is a mash-up of the German word for “rim” (Felge) and “GAN”, the latter being an acronym for Generative Adversarial Networks. GANs are a special form of self-learning computer program in which two algorithms compete as opponents during the so-called training, becoming better and better in competition with each other.
It works like this: One of the two algorithms, the “generator”, makes artificial images of a specific motif – in the case of FelGAN, a vehicle rim. The discriminator – the competitor, so to speak – sees a selection of images, consisting of real wheel photos alongside images from the generator. Now the discriminator decides whether each image is a creation of the generator or a real photo. This process is repeated again and again until training is completed. Both algorithms are designed to learn from their mistakes and improve continuously. After enough runs, the generator’s creations are so deceptively real that even the human eye cannot, or can only barely, distinguish them from real photos.
The application’s intuitive user interface, which is based on Streamlit technology, creates short development cycles and quick feedback between the design and IT team. So that designers do not have to rely on high-performance local hardware when using the software solution, the components of the AI application – which require a lot of processing power – are run in the cloud.