Lyra Martin Tech 06.27.2022 11:Jun AM EDT

DALL-E Mini Internet's Favorite AI Meme Generator Now

Over the last several weeks, DALL-e Mini has been producing around 50,000 photos every day as it became the internet's meme generator.

Despite that, DALL-e was not trained on explicit content and is designed to limit certain phrases. Some of those risks can become more difficult to manage over time.

How Did DALL-e Mini Gain Popularity?

With some recent updates and a few viral tweets, DALLE Mini's capacity to clumsily draw any kinds of weird, amusing, and even nightmare images quickly became meme magic. Check out its recreations of 'Thanos searching for his mother at Walmart,' 'drunk naked males strolling around Mordor,' 'CCTV camera video of Darth Vader breakdancing,' and 'a hamster Godzilla in a sombrero invading Tokyo.'

Open source artificial intelligence platform Hugging Face witnessed a surge in visits to this AI image-generation tool, DALL-e Mini. It's been over a year since an independent developer created this basic application. It uses a single line of text to produce nine graphics in response.

Hugging Face's servers were overloaded as more individuals uploaded and posted DALL-E Mini photos to Twitter and Reddit and as more people joined the service. DALL-e Mini has been delivering around 50,000 images every day in the last several weeks.

The visual style of DALL-e Mini is very foreign. Frequently, things are warped and smeared, and individuals have faces or other body parts missing or deformed. However, you can generally see what it's trying to show, and it may be entertaining to contrast the AI's sometimes irrational output with the initial query.

DALL-e Mini Risks Are Hardly Manageable

DALL-e was another example of how advances in AI research are often promptly copied elsewhere, sometimes within months. The initial DALL-e research article captivated Houston, Texas-based machine learning specialist Boris Dayma. He put together the initial iteration of DALL-e Mini during a hackathon hosted by Hugging Face and Google in July 2021 despite OpenAI not disclosing any code.

Dayma has recently improved the first version, which often generated blurry, low-quality photographs. He changed the name of his project last week to Craiyon at OpenAI's request to prevent confusion with the original DALL-e project. A paid version of Dayma's picture generator is also being planned, and the new website includes advertisements.

The AI model behind DALL-e Mini develops images by identifying correlations between words and pixels from 30 million tagged photographs. Dayma's training data came from several public image collections, including one from OpenAI. The system makes mistakes because it doesn't know how the real world works. Small text snippets are generally ambiguous, and AI models don't grasp them. Still, Dayma has been amazed by the results of his recent efforts. The "Eiffel Tower on the moon" was his most creative prompt. Now, crazy behavior works.

But some of those creative exercises led DALL-e Mini to strange places. The technology is intended to restrict specific terms and was not trained on explicit material. However, individuals have posted pictures from questions related to the World Trade Center attack, school shootings, and war atrocities.

Some of those dangers may become harder to control. Dayma, the designer of DALL-e Mini, says it's just a matter of time until tools like his can create lifelike graphics. But he believes AI-made memes may have prepared humans for that scenario. He says, "You know. I hope DALL-e Mini teaches people that images aren't always true."