What will happen when an AI is trained with tens of thousands of album covers from Spotify playlists over the course of several days?
With a bunch of Python scripts and Generative Adversarial Networks (GANs), a Spanish artist named Bruno López, managed to create a trippy music video.
What López did, was first downloading every track featured on Spotify’s official editorial playlists, which consisted of 50,000 albums covers with the resolution of 640x640. López then trained the AI using Google Colab Pro using skyflynil’s implementation of the original StyleGAN2, which is the same system that previously created the fursonas that don't exist, among others.
After the training was done, López created a Python script to organize the results by the colors and faces present in the covers.
He then animated the images to the beat of the Sound Stabs song “Interference.”
According to López, the Python script for the music video is based on lhzbrian’s, which created random latent vectors that are used for the network to generate the images, which were later interpolated to create an organic, AI-generated, animated transition between all of them.
To synchronize the animation with the song, López adjusted the speed of the interpolation by directly linking it the volume of the beat on each frame using PyDub.
To make the end result more diverse, multiple clips were generated with different attributes each and later put together in Adobe Premiere, resulting in the “Interference” music video.
While the music video consists of some flashing images, which can be uncomfortable for some viewers, the AI is seen capable of creating an aesthetically pleasing video that matches the beats.
What's more, the AI is capable of generating its own album covers.
This kind of AI is nothing new.
In the past, for example, an AI was able to generate music by seeing silent piano performances.
What's more, there is also an AI-powered music generator that loves Pop music, an AI that loves and also capable of generating Heavy Metal music, one AI that is capable of generating the "perfect TikTok" song, and many others.
And more recently, there is an AI that can visualize poem, in a disturbing way.
López here, doesn't introduce anything much to the table that can swerve or point towards a path for more powerful AIs.
But his project is an example that shows AI as an intelligent machine.
López' project shows that machines can indeed become smarter by doing what is out-of-the-box, by just extracting the patterns from the data it learned.
By having more capable platforms and more powerful models, results are only restricted by the number and the quality of the data.
This is why the scope and the breadth of AI is limitless, and it's only up to one's imagination to decide.