Playing with AI-system DALL-E

Get started with DALL-E, the AI-system from OpenAI that composes new images from a daily language sentence.

Playing with AI-system DALL-E
“A red automobile painted by cezanne ”

For quite some time I saw the synthetic images of DALL-E popping up, and I read a lot about it. It was only accessible to invited guests, but now anyone can play with it. And so, I played with this fascinating technology.

The name DALL-E

DALL-E is the name given by the company OpenAI to their AI system. The name DALL-E is a portmanteau of the names of Pixar character WALL-E and Spanish surrealist artist Salvador Dalí. A well-chosen name for the often surreal images created by the computer.

Wall-e and Dalí on DALL-E.

What is DALL-E?

For the user, it is software that lets you use a textual clue (prompt) or existing image to request a new image. This is then not searched, but created. A new, synthetic image.

Making those images was not the big breakthrough, but the fact that by giving a textual command, you get an image. The engine behind this is GPT-3, a so-called Large Language Model. A model for everyday language processing based on massive sets of information.

So first, the given text in natural language will be interpreted by GPT-3 and broken into bits of meaning. Then, using those words, the system creates a new image from snippets it has scraped from the internet. So actually, the images are a sort of collages.

An example of playing with DALL-E

The underlying technology is difficult to understand in detail, but you can get an intuitive feel for it. Especially if you get to work with it yourself, you will learn a lot.

To get started, go to OpenAI's website and first create an account:

https://openai.com/dall-e-2/

And basically, you can then get started right away. By entering an English text or your photo. Here is the example I work with:

A childs drawing of a happy dog singing the blues

That got me these four funny children's drawings. The dog even has a blue tongue on two of them:

A childs drawing of a happy dog singing the blues

Not bad for a first attempt. You are shown four variations each time and by choosing one image you can request further variations. On the left, the original and then the four images from the second round.

Variations of: A childs drawing of a happy dog singing the blues

The programme provides you with ample tips. Ask for 3D renders, or add your favourite painter or style. For example, Japanese woodblock art ukiyo-e.

Peacocks and blossom in the style of ukiyo-e.

Playtime over

But not immediately game over. You get a few free credits every month, but you can buy extra if needed. They cost $15 per 115 credits at the time of writing.

I must honestly say that I could play for hours and some images go straight into my dreams. So scary and realistic. It's absolutely fascinating, and it's a game with language too. What can I describe in a way that something comes out that I wanted or that surprises me. Some assignments give extraordinary results, and then others are mediocre.

What struck me is that all the results are square, which is not always what I require (why you need portrait and landscape). But actually, many images are perfectly usable. There are, of course, questions about copyright. There are also ethical questions about deepfakes, for example. Some images are deceptively real, like the hawk on the handlebars of a bicycle. DALL-E provides a spectacular new technology that will have influence that stretches far beyond the digital world.

Red-tailed hawk on the handlebars of a bicycle.