Developed Disco GAN — an experimental generative model addressing the challenge of producing high-fidelity visual output from a very small, highly specific dataset. The project explored whether a neural network could learn to capture and "distill" the intangible atmosphere of a specific Dutch discotheque into visual art. I designed and optimized a hybrid training pipeline combining VAEs and GANs on HPC infrastructure (SLURM/Docker), iterating on architecture choices and data augmentation strategies to extract meaningful representations from severely limited training data. This work was selected for presentation at the Betweter Festival 2023, sponsored by SETUP. Here an extensive article can be found on the topic.