Artist Uses AI to Recreate ‘Watermelon Sugar’ Music Video Using Footage of Ronald Reagan

Ronald Reagan mashes well with the Harry Styles song Watermelon Sugar. I know this because someone has designed an AI with the exclusive task of recreating the Watermelon Sugar music video using footage pulled from the Reagan Presidential Library.

André Röhrig, a German living in the UK and working in data science, is responsible for the remixed music video. “My AI automatically detects the scenes in the source video clip and then replaces them with similar looking archival footage. This process is end-to-end and doesn't require any manual intervention,” the YouTube description reads. “The archival footage comes from the Reagan Library.”

The AI did a passable job of recreating Styles’ pop song using footage pulled from the Reagan Archive. Both videos open on an ocean scene. When the original video cuts to footage of Styles’ sitting at a table covered in an orange tablecloth, Röhrig’s remix cuts footage of Reagan at an orange table reading paperwork. The AI approximates an imitation based on what it has access to. It’s not perfect, but according to Röhrig, that’s the point.

“These music videos have been an attempt to force interesting visual errors, i.e. when the used archival footage is close and very far away from what it's supposed to be at the same time (e.g. when Stalin appears in a Lana Del Rey video),” he said. 

Röhrig did something similar with the Muse song Uno and Lana Del Rey’s Video Games on his YouTube channel, priorai. These other videos used footage from the Reagan Library, but the Lana Del Rey video also has footage from what appears to be Soviet sources. There’s black and white video of Soviet soldiers as well as a brief and flickering reel of Joseph Stalin. “The Reagan Library contains some anti-communist propaganda documentaries, and that's where the footage about Stalin, Hungary 1956 etc. comes from,” Röhrig told Motherboard in an email.

Röhrig said he got the idea for the project when he noticed that AI and machine learning often fail in interesting ways. Artist Mario Klingemann is also an inspiration, he said

The original music videos for both Watermelon Sugar and Video Games are nostalgia trips that copy the aesthetic style of the 1960s. According to Röhrig, this was important to helping his AI—which he said he designed and trained himself—understand how to make its own music video. 

“There are certain parameters like hard cuts and short scenes that are important for the process not to fail,” he said. “And the video footage the AI tries to replicate must be represented to a certain extent within the distribution of the archival footage, otherwise the results are not very satisfying. This is the case for these retro style videos I've used here; the 'real' Lana Del Rey video even uses lots of archival footage.”

According to Röhrig, creating the AI took about two weeks. Why the Reagan Library, though? After all, there are plenty of other sources of workable video, including the libraries of other American presidents. Röhrig said that it produced the best results. 

“Initially I was looking for footage in the public domain for practical reasons, so the videos don't get blocked because of copyright claims,” he said. “This is the case for these public domain presidential video collections. I've tried other collections (Trump, Obama), and also other sources like old educational films produced by the U.S. Government, but so far I've had the best results with the Reagan Library footage. The more recent presidential videos have much better production videos and more editing, while the Reagan videos are often just someone filming random people at campaign rallies for hours, which is great for a diverse data set vs. just lots of shots of the president.”

Other videos on Röhrig’s channel show an AI that’s trained to make abstract art using a Generative Adversarial Network (GAN), a common technique for creating artificial images.

“This animation of abstract images was created by an AI that was trained without any training data,” Röhrig said in the video’s description. “Instead, an adversarial training process of different deep convolutional neural networks against each other was used to create 'pure' abstract visuals.”