AI Spits Out Nightmarish Images Based On BuzzFeed Headlines
BuzzFeed articles are crafted for the everyday reader, but whether or not this extends to artificial intelligence is still up in the air.
When you get past the convoluted symbols of X Æ A-12 Musk’s name, you’ll probably just envision a cute baby. AI’s understanding of how Musk babies are made, on the other hand, involves the merging of Elon Musk and Grimes’ facial features (well, it’s not entirely wrong), plus a mullet.
This fever dream began when the work of generative artist SOMNAI, real name Max Ingham, was spotted by BuzzFeed data scientist Max Woolf. SOMNAI experiments with a relatively new neural network called Contrastive Language-Image Pre-Training (or CLIP), and Woolf had a niggling curiosity about how its effects would play out on BuzzFeed headlines.
“Good idea,” the artist responded. “Gonna load some up now.”
Thanks to this brief online exchange, we now have the machine-learning manifestations of titles “How I Cooked 20 Pounds of Cabbage,” “People Are Freaking Out Over This Pink Mac And Cheese That Kraft Is Releasing For Valentine’s Day,” and “Grimes Just Gave Her And Elon Musk’s Baby A Haircut, And The Result Is Very… Interesting.”
`How I Cooked 20 Pounds of Cabbage`
— SOMNAI (@Somnai_dreams) February 7, 2022
`People Are Freaking Out Over This Pink Mac And Cheese That Kraft Is Releasing For Valentine’s Day`
`Grimes Just Gave Her And Elon Musk’s Baby A Haircut, And The Result Is Very…Interesting`
(@BuzzFeed guided Diffusion) https://t.co/xGHgL60YtI pic.twitter.com/AgHYWNLOyD
Images by SOMNAI and featured with permission
The result is “very interesting” indeed.
As reported by Futurism, who first spotted the project, these visualizations are particularly interesting because they’re powered by a diffusion model, which is also a fairly new concept. It achieves its effects by corrupting the training dataset until the data is no longer recognizable, and then getting the neural network to “reverse” this corruption. The end product is often terrifying.
The open-source CLIP neural network that Ingham uses is particularly clever at picking up visual concepts from natural language, so we’ll just put these results down to a taste for the macabre.
You can follow the artist here to view more of his work. They’re not all this eerie—and are mostly quite beautiful, in fact.
Diffusion animation anyone? pic.twitter.com/lro71D1tYL
— SOMNAI (@Somnai_dreams) January 1, 2022
[via Futurism, images by SOMNAI and featured with permission]
