DALL.E 2: between Dalí and WALL.E, an AI capable of creating images from random text

Benjamin Logerot

April 11, 2022 at 4:56 p.m.


The artificial intelligence research company OpenAI presented the new version of its image generator via keywords: DALL.E 2. A more precise and comprehensive generator that allows to create infinite combinations of artistic images in high resolution thanks to the power of its AI tool.

However, DALL.E 2 is currently reserved for groups of researchers until the AI ​​is properly developed and the risks of misuse are eliminated.

One text, one image

In the dedicated field, mark “ An astronaut is lounging in a tropical hotel in space in pixel art style ” and hop ! You get an image with precisely everything you wrote. Not a fan of the end result? Change “pixel art” to “Van Gogh” or “astronaut” to “a dog” and the image appears modified with the new result. This is what DALL.E 2, from the OpenAI company, offers.


© Open AI

Launched in January 2021, the DALL.E tool (a mixture of Salvador Dalí for the artistic side and WALL-E the little robot for the technological side) already made it possible to do more or less the same thing but, a year of development later, the researchers behind the project are able to release an even more advanced and more accomplished version.

Thanks to this tool not yet available to the general public, three things are possible:
create images from keywords; create variations of already existing images (take the Mona Lisa and suggest, for example, dressing it up with an Iroquois haircut) or merging two images together.

Artificial intelligence at the heart

In concrete terms, how does that work ? The tool’s site explains it pretty well. DALL.E 2 uses a neural system trained with images and their description. Deep learning allows the tool to understand which word belongs to which image by analyzing and cross-checking the patterns of thousands of photos associated with a given word. For example, for the word “koala”, the tool will have previously explored the database of millions of photos to define what a koala is.

When creating the image, the tool uses a so-called “diffusion” process. Starting from a pattern of randomly placed dots, it gradually changes to an image as it recognizes specific aspects of that image. Obviously, there are limits. If an airplane has the word “car” in its description, when you want to create a car it can put the image of an airplane since for the AI, a car is then an airplane.

DALL.E 2 is a research project and therefore not available on the API. A selected group of users participate in the research and only trusted researchers can register to participate. Work on the security of the tool is carried out in order to prevent the generation of violent, hateful, political or pornographic images. It is also prohibited to create images using photos of real people or personalities.

Source : Smithsonian Magazine

Source link -99