Have you ever wondered what Homer Simpson would look like if painted by Vincent van Gogh? What about Batman riding a skateboard?
Well, you don’t have to wonder much longer. A hugely popular artificial intelligence art generator has taken the internet by storm, turning curious users’ most outlandish thoughts into reality.
The software, called Dall-E mini, is a free, open-source AI that produces images using text prompts.
Users just need to enter a simple description and hit ‘run’. Let’s use “Batman riding skateboard” as an example.
In just seconds, the program can interpret the description and spit out nine images that match the request.
And voilà, watch Bruce Wayne transform into Tony Hawk before your very eyes. Here are some creations The new daily newspaper previously assembled.
How does it work?
Named after Spanish artist Salvador Dali and Disney Pixar robot Wall-E, Dall-E Mini is the brainchild of Houston-based programmer Boris Dayma.
The model really started making waves online in the past two weeks, but Mr. Dayma first built the program in July 2021 as part of a Google AI competition.
He says Dall-E Mini takes anywhere from 400 to 500 million pieces of “unfiltered data from the Internet” and puts them together to fulfill users’ requests.
Speak with the me, Dayma said the driving idea behind the creation was to make AI accessible to ordinary people.
“It was both a technical challenge and an interest in having something publicly available.”
Work in progress
The program can be a bit disappointing – which Mr. Dayma openly admits.
Dayma says the app sometimes struggles with more precise details, such as faces.
“The hardest thing is definitely the people,” he said. “When you draw a landscape with Dall-E, it’s great because if there’s a small problem with a tree, nobody notices and the landscape still looks great.
“But if there’s a problem with a face, we notice it. If there is a little flaw with one eye, we can see it. With an avocado, even if it has flaws, it’s good enough.”
But it’s all part of the process. Like most artificial intelligence, Mr. Dayma says Dall-E mini learns as it goes.
“The model is still in training. It’s still going to improve. Day after day it only improves a little bit, but week after week you really notice it.”
While he worked painstakingly to train the model in the early days, it is… now training blind – learning because it generates numerous images based on the requests of its users.
Mr. Dayma has documented the program’s lessons over time. This tweet perfectly illustrates the evolution of AI’s capabilities. Using the prompt ‘A Pikachu armchair’ as an example, we can see how far Dall-E mini has come in just a few days.
As can be seen in the top row, the interpretation of the model is incredibly ambiguous. But as time goes on, it gets much better to merge the two concepts.
Dall-E’s bizarre mini-creations have spread like wildfire online, some with hundreds of thousands of likes each. Here are some of our favorites so far.
Want to make your own Dall-E mini creation? Click here.
More AI news: