album art generator made using GANs and deep learning

this was my (and a partner's) final project for eecs 442 at university of michigan. it was focused on creating album covers that were particularly similar stylistically to others within a given genre label

the focus of this project was to use generative adverserial networks (GANs), transformer architectures, encoders, and decoders. these were, at the time, state-of-the-art deep learning techniques used to create convincing generations of many modalities, including audio, visuals, text, and more. these models can be seen as a sort of precursor to the following generative ai architectures seen in the early to mid 2020s on chatgpt, deepseek, claude, and more

the workflow was that of a typical machine learning project. my project partner and i handled all stages of the process, from data collection and engineering, all the way to hyperparameter and neural net architecture tuning. this process was iterated upon multiple times as we compared the results

in the end, the results were admittedly not too exciting. only by the end were any real forms or figures emerging. individual images seem like they had been painted, then smeared or smudged before having the chance to fully dry. made for some interesting images though

i have to admit, however, that the results did scare me a little bit too. it's funny to say that since you can look straight above this paragraph, study the best results we achieved, and then be very justified in saying these did not resemble anything of any substance whatsoever. but still, the fact i could just generate arbitrary, odd content this simply was bizarre

it was a little sobering considering that, for a deep learning model to function effectively, you need exponentially more data points that we dug up. not only that, but they also needed to be of much higher quality than what our data source yielded. essentially, we starved this model architecture of all nutrition and substance to learn effectively from, and returned these images that very much had character, form, and structure to them. it's usually garbage in garbage out, but something about these outputs did not feel like garbage to me. one very strongly reminds me of gorillaz's up on melancholy hill, which threw me for a loop. very bizarre. in a way, the limitations of our project showed me a glimpse into the future of generative ai

if there is any interest in reading our presentation abstract, it can be downloaded right here