This AI image generator lets you type in words and get weird images back.

 

It took 19-year-old Matt Leaming from the UK just a month to reach one million followers on Twitter. And all he needed to do was share a steady stream of the most exotic computer-generated images he and a bunch of internet strangers could think of.

In recent weeks, the digital marketing apprentice, who goes by the name @weirddalle online, has shared photos of people emptying into the woods, the Demogorgon from Netflix's "Stranger Things," a basketball. is holding, and a Beanie Baby that looks like Danny. De Veto.

These and other images, which range from funny to disturbing, were created with a freely available artificial intelligence system called Craiyon. To use it, you simply type in what you want to imagine — "a rainbow tiger eating a slice of pizza" — and it will spit out images in response.

The brainchild of Boris Dyma, a Houston-based machine learning engineer, Craven is popularizing the growing trend in AI. Computers are getting better and better at digesting words and creating increasingly realistic-looking images in response. Recently, people are typing about 5 million prompts per day, Daimah said.

There are similar, much more powerful AI systems than Craiyon, such as OpenAI's DALL-E (Craiyon was initially named DALL-E Mini) and DALL-E 2, as well as Google's Imagine. But unlike Craiyon, which anyone can try, most of them are not available to the public: DALL-E 2 is open to users by invitation only, while Imagen is not to users outside of Google. has been opened.

"I think it's important to be able to have an alternative where everyone has equal access to this kind of technology," Daima said.

However, in the process, Craiyon is effectively serving as a test of what might happen - good or bad - in the future if someone could gain access to such AI systems and interact with them with just a few words. Can ask for any type of image. And as with many new technologies, it's a work in progress; In the near term, if left unchecked, this can lead to consequences that reinforce stereotypes and prejudices.

The Infamous BFG

Dema and a few other coders built the AI ​​system last July during a hackathon sponsored by Google and Hugging Face, a company that builds and hosts AI models. Initially, Daima said, he built it as a technical challenge. He thought DALL-E was cool and wanted to make one himself. He posted the text-to-image generator—then called DALL-E Mini—on Hugging Face, where anyone could try it out (it's still available there under that name). But it didn't get much attention beyond the AI ​​community until the last couple of months, perhaps because of the limited quality of the images it produces.

In the past, for example, it would be able to visualize simple things like landscaping, Daimah said. But slowly, he's done things like fix bugs and improve the code, enabling him to get better at coming up with more complex images, like the Eiffel Tower landing on the moon.

"I was very happy when the model started drawing it," he said. "But then people came up with more creative things, and somehow the model got to a point where he was able to do something that looked like he was asking for, and I think that was a turning point. "

The images Craiyon creates aren't nearly as realistic as the DALL-E 2 or Imagine can come up with, but they're still appealing: people blur things out, and images look fuzzy and minimal.

For now, Craiyon is mostly used by people like Laming for fun - probably because the results aren't nearly as crisp or photo-realistic as what you get from the DALL-E 2 or Imagen. Can, but also because people are still trying. Find out what to do with it. (The Craiyon website currently runs ads to offset the costs of the servers that power the AI ​​system, and Dayma said they're trying to figure out how to make money off of it while letting people use it for free. should also be allowed to play.)

 

Many of Lamming's images on Twitter come from a Reddit forum he created for people to post tips and the resulting images when they were run through the system, which is how he used one of his own. Opts for another Twitter account, which he runs, @spotifyweird, which tweets weird Spotify playlists.

Lemming's most popular tweet of all time was a June 14 post with the prompt "Fisher Price Guillotine," which was initially posted on his subreddit by a Reddit user. Popular posts can take something from the news or pop culture and combine it with something completely random or shocking or gross — like Minions-themed urinals — or just come up with a funny play on words. (Think "The Notorious BFG" or "Ice Cube in an Ice Cube").

As users become more familiar with Craiyon's types of results, the prompts become increasingly specific to the types they want to see — such as calling for a medical illustration of a burrito or courtroom sketches. Showing how it can be. If a capybara sued Elon Musk. Sometimes they're really weird, like in this photo of archaeologists discovering a plastic chair.

To come up with a good prompt, Lemming suggests, just "think of the most unusual situation to put someone or something into." In fact, the cues that lead to these images are a new form of creativity themselves.

Biases on display

Mar Hicks, an associate professor at the Illinois Institute of Technology who studies the history of technology, said the AI ​​system reminded him of early chatbots like Eliza, which MIT professor Joseph Weisenbaum developed in the 1960s. was created and intended to imitate a physician. . Such programs could convince people that they were communicating with another human, even though the computer did not truly understand what it was being told (Eliza gave scripted responses).

"I think it's appealing in the same way that a game of chance appeals, or a party game," Hicks said. "There is uncertainty about where something is going to happen."

But Hicks is concerned about the AI ​​system's ability to respond to any written prompts with images, rather than the occasional error message indicating that it doesn't know how to type a person. How to parse a given sentence "That means you're going to be taking out the garbage occasionally," he said, and the onus is on the consumer to find out why. This was the case with some of the gestures I fed the crayon, making it occasionally frustrating and frustrating to use, but Daima pointed out that it wasn't easy to predict what it might draw. Or not, and sometimes the results are surprising, or at least surprising. weird

Daima said she's heard from people using crayons as logos for new businesses and as images in videos. (OpenAI and Google have suggested that their systems could eventually be used for things like image editing and creating stock images.)

 

While there may be creative potential for these AI systems, they share a key problem that permeates the AI ​​industry at large: bias. They're all trained on data that includes a wide variety of internet content, which means the images they create can also reveal a number of biases, including gender, racial and social stereotypes.

Such biases are also evident in Crayon's ambiguous-looking images. And since anyone can type anything they want into it, it can be a disturbing window into how stereotypes can creep into AI. For example, I recently prompted Crayon for "lawyer," and the results were all blurry images of what appeared to be men in black judge's suits. Meanwhile, "A Teacher" immediately finds the only figures that appear to be women, each in a button-down shirt.

Daimaa is aware of this. A "Frequently Asked Questions" section on Craiyon's website states that the model's reliance on internet data results in "images that contain harmful stereotypes", and that Craiyon's biases are behind this. are working to document and analyze Daima noted that many AI systems are biased, whether users are aware of it or not, and said he likes that everyone can see Craven's biases directly in the images he creates.

He also said that he tried to prevent the model behind the crayons from learning certain concepts to begin with. However, it only took a few minutes for me to spot some obvious clues that led to images that, to put it bluntly, were not safe for work.

When asked if he thought its general availability might be a bad thing, given its obvious biases, he pointed out that the images that come out with it, while looking better than in the past , are clearly not realistic.

"If I build the Eiffel Tower on the moon, I hope no one believes that the Eiffel Tower is really on the moon," he said.