A new AI text generator is bringing an extra dose of absurdity to the internet’s favorite meme formats.
Developed by Imgflip founder Dylan Wenzlau, the web app uses natural language processing to generate wholly original text for any of four dozen meme templates in a matter of seconds.
The output mimics the internet slang and jumbled in-jokes of the nearly 1 million meme captions on which the system was trained, with mostly nonsensical–often hilariously so–but sometimes surprisingly coherent results.
Check out some of the examples we generated below:
The generator is the latest example of how recent advances in natural language AI have helped enable a new era of creativity in machine learning, fueling everything from sophisticated chatbots to cult favorite text adventure games. However, unlike many of these projects, Wenzlau opted to generate text at the character level, rather than word by word, allowing the output to match the disjointed syntax that characterize most memes.
“Character-level generation rather than word-level was chosen here because memes tend to use spelling and grammar … uh… creatively,” Wenzlau said in a Medium post that thoroughly documented the process.
The captions that make up the training data are all pulled from a set of around 100 million memes hosted on Imgflip, which were then culled down to 20,000 each of the 48 most popular formats for a total of 960,000. Wenzlau also posted the code on Github for others to replicate his work.
The app is titled, “This Meme Does Not Exist,” in reference to a growing trend of novelty social media accounts and online apps offering various AI-generated ephemera, from garbled AI images of cats—“This Cat Does Not Exist“—to anime characters. Projects like these have been enabled by very recent breakthroughs in both language and image generation that have spawned a nascent AI art and creativity scene.