Infinite Bad Guy is an AI experiment by YouTube and Google Creative Lab, launched to celebrate Billie Eilish's "Bad Guy" hitting 1 billion views. The experience aligned over 15,000 fan-made covers, spanning genres from death metal to rubber chicken, shot everywhere from bedrooms to suburban exteriors to fully choreographed studio sets, into a seamless unique music video every time you watch it. Watching every possible combination would take 1.46 × 10100 years.
The project won the 2021 Webby Award for Websites and Mobile Sites, Music.
Trained an early variant of CinemaNet using the Keras framework. This model was used to power the algorithms that chunked up user generated covers into atomic blocks that could be put together seamlessly for the end user experience.
At the time, ImageNet based models were the state-of-the-art for generating image embeddings, but these simply didn't transfer well to this domain. CinemaNet worked much better as it understood visual nuances of cinematography and produced better embeddings and tags. Also curated early versions of the custom datasets that were used to train the model.
Learn more about the technical details in this blog post.