► What is Dall-E?
Dall-E is a computer program, doped with artificial intelligence, which creates any image from a text. Just ask it “a cat in a flying saucer in the style of a medieval tapestry” and the machine displays four drawings representing exactly that. Or “a corgi with disco glasses in the style of Andy Warhol”, and you will see a very pop art dog appear.
First reserved for AI researchers, then for a few thousand people, the program opened to the general public this week. All you have to do is create an account and enter your phone number to access the interface. With a limit however: each generation of images from a text uses a credit. The user has 50 free credits upon registration, then 15 credits per month. These credits are not cumulative from one month to another, but it is possible to buy additional credits.
► How does it work?
Dall-E is an artificial intelligence to which the teams of the OpenAI company have provided more than 650 million images with their descriptive texts. So no one taught the AI to draw, in the classic sense. She simply has an enormous visual culture, a bit like a forger might create an unpublished painting by Gauguin. Its source code, the cooking recipe that explains how it works, has not been disclosed but is based on GPT-3, the company’s other project.
This first project was not about images but text. By supplying instructions or even a beginning of text to GPT-3, the program writes a corresponding text or the continuation. Perfectly stunning for creating product descriptions but also news articles – the British newspaper The Guardian had an article written by AI in September 2020 – GPT-3 had aroused much criticism and concern about the dissemination of false information and for its biases, for example associating Muslims with negative adjectives.
Other programs using neural networks had already raised such questions. The deepfake, where videos of very real people are manipulated, or the site “This Person Does Not Exist” (“this person does not exist”, in English) have upset the relationship to the truth of images.
► Can we have everything drawn?
No. The application does not provide a “creation” for violent, racist, hateful, sexual etc. texts. Also impossible to generate representations of politicians or celebrities. The query “Queen Elizabeth II on a flying dragon” comes to nothing. And trying to force the hand of the virtual painter exposes you to account suspension.
Other than these restrictions, it’s hard to find text that Dall-E doesn’t convert to pixels. But do not imagine that artificial intelligence immediately provides incredible paintings. The result depends a lot on the text provided, which must be in English for now. Unsurprisingly, the more complete it is, the more faithful the rendering. The AI works by keyword, so it is not necessary to provide it with real constructed sentences. Problem: she can mix up the order. Asking him for “a purple glass and a red fork” can sometimes result in a red glass and a purple fork.
Between curiosity and concern, many artists wonder about this tool. Some are calling for new copyright and attribution regulations. Others recognize that they use Dall-E to generate images to copy, and work, for example, on the representation of the muscles of a horse in full gallop. Faced with the vast field of creativity that is opening up, Dall-E has not finished upsetting the relationship to art.