Have you been wondering what the hilariously chaotic grids of images circulating social media really are? So have we. They’re made by software called DALL-E mini and this is what we know.
What is DALL-E mini?
DALL-E mini is an artificial intelligence program which allows users to type in anything (really, anything) and the software will automatically generate original images from the text.
It’s based on DALL-E, a “12 billion parameter” neural network first written about in January, which does the same thing but, from what we can make out, a bit (lot) better.
The original paper, published on Open AI, gave some pretty accomplished examples: an avocado the shape of a chair, or a pikachu purse.
DALL-E mini takes the training wheels off, and lets users unleash the full creative horrors of their inner subconscious. And that’s where the fun starts.
Nicholas Cage in parliament? Sure. Jar Jar Binks at the Nuremberg trials? Make it so. Understandably, people have been excited to share their surrealist creations on social media.
What is DALL-E used for?
If you want to see Dumbledore eating a Big Mac, pop it into the generator and voila – you have “Albus McLovin it”. Apologies to Harry Potter fans everywhere.
Why is it called DALL-E?
The AI was named after Pixar’s Wall-E, the doe-eyed robot that cleans consumerist waste on a post-human planet. The name also nods to surrealist painter Salvador Dalí.
Here’s DALL-E mini’s attempt at putting WALL-E in a Dalí painting. Masterful.
Is DALL-E public?
Yes, the generator is accessible to the public. Unsurprisingly it has not taken long for the online world to start sharing their most hilarious and avant-garde artworks.
Here are some of our own creations.
Boris Johnson on Strictly Come Dancing. We like to call this “BoJo’s Tango No-no”.
Beyonce getting that Asda Price feeling.
The right honourable gentleman, Nicholas Cage, the new secretary of state for nightmares.
Early career pics of DJ Iron Lady on the ones-and-twos.
In another universe, Jeremy Corbyn was a tank commander. Still had the same suit though.
How does AI generate images?
The program consumes billions of images from the internet along with their descriptions and builds associations between the two. The idea is that if you feed it enough images of, say, a tree, and those images are labelled “tree”, it will then be able to identify images of trees that don’t have labels.
Remember all those times you’ve had to identify a car or traffic light to prove you’re not a robot? Well congratulations, you’ve been teaching computers to see.
Originally used to create the captions for images, the AI can now do the opposite, developing images from captions.
As the AI learns from the unfiltered data of the digital universe it’s still limited by “societal biases” and runs the risk of exacerbating “stereotypes against minority groups”, the generator’s disclaimer states.
Who is behind DALL-E Mini?
DALL-E mini is currently operating on the website Hugging Face, described as an “AI community” aiming to “democratise good machine learning.”
The generator is based on a programme by artificial intelligence research facility, OpenAI, that have worked on numerous projects including RoboSumo, which are virtual humanoids, or Debate Game, teaching machines to debate problems. OpenAI aims to create “general intelligence [that] benefits all of humanity”.
How long does DALL-E take?
The generator takes between 15 and 120 seconds to cultivate nine images, but it depends on the amount of users trying to use the generator. A lot of traffic may mean it won’t run immediately.