They have been flourishing on the Internet for several months: artificial intelligences such as Dall-E 2, Midjourney or Stable Diffusion in a few words create new images, each more impressive or strange than the other. A fast-growing sector that also attracts tech giants.
You wake up from a strange dream that you no longer remember… except for the surreal image: Joe Biden hosting Mickey Mouse at the White House. How to immortalize a slightly blurry scene without erasing it from your memory?
You don’t need to master Photoshop or even touch a pen before: today, thanks to artificial intelligence, anyone can create photorealistic scenes, realistic works of art or images from a fever dream in less than 30 seconds. Here is our American scene, born of an imperfect, strange, but simple sentence.
To do this, simply visit one of the recently opened public sites, such as Dall-E. No frills in the interface, just a text bar. Here’s all you need to create the scene of your dreams: first, as impossible as it may be, describe in a few words (and in English) the image that needs to be created.
For example, “A golden retriever on a surfboard in space, riding a nebulae”. Then click “Create”, wait about fifteen seconds… And there you have it, the AI has created an image out of nowhere.
The image is not a copy of a sketch already available on the Internet, but a unique result of a completely artificial creation imagined in real time.
As proof, these programs are far from perfect. Some of the productions are even downright disturbing at times, or downright unsuccessful—especially the often distorted human figures.
To get a completely coherent image, you need to go over it several times and refine your query by adding phrases: style (“polaroid”), light (“cinematic lighting”), expected level of detail (“high detail”), artist names (” Picasso”)…
Here’s our golden retriever on the surfboard again, this time showing specifics: an oil painting version with lots of detail.
New scenery for iconic paintings
These artificial intelligences are not only used to create photorealistic situations or works of art almost out of nothing. They can also enlarge images.
Have you ever wondered what goes on around famous painting designs? The girl with the pearl? It’s very simple: once again copy and paste the original image into Dall-E, then click on the area where you want to expand the image and type the desired result.
Again, it often takes a few tries, but the results can be amazing, like this achievement by American artist August Camp.
And here it is, Tech&Co produced it in about 30 minutes – with all the flaws of proportion and style – with the same tool, based on the artwork Milk girl By Johannes Vermeer.
AIs can also reprocess entire images, for example turning simple children’s drawings into professional work. Or give your selfies a radically different style by modifying them… Just typing a few words at a time, just like a web search.
“It’s not a simple Google search: it creates a completely new image that doesn’t exist anywhere else,” insists Valentin Schmitt, a professor at Sciences Po and author. About Art and Artificial Intelligence.
“Machine learning” and “hidden space”
How do these AIs keep up? Programs must “learn” to decode real images before they can create new ones. To do this, researchers use the hard way: they are forced to swallow hundreds of millions, if not billions, of all kinds of images recovered from the Internet and accompanied by written descriptions.
These programs are then trained to detect repetitions in images using “machine learning” techniques that allow the program to improve itself almost autonomously. Thus, artificial intelligence will learn by itself to distinguish a dog from a cat, a picture from a painting… Then each picture is stored in a kind of large virtual storage – an area called “hidden space”.
That’s it for the “inspiration” part. But when you ask these AIs to create an image, they first need to understand what you’re writing. This is possible thanks to another module that studies written representations to learn which part of the image each word corresponds to, thereby understanding natural language.
After deciphering your request, the AI will determine which part of its secret area or which shelf of its warehouse is most suitable for it. Then he will start creating the image. Unlike a human artist, it doesn’t start with a blank page, on the contrary: it starts with a collection of pixels of random colors, “cleans” them little by little, changing some to get closer to the required image. . A technique called “diffusion”.
After pictures, music, podcasts…
Artificial intelligences that can generate images on demand have been developing rapidly recently. “We’ve seen a real explosion of this technology in the last 3-4 months,” says Valentin Schmitt. In fact, these tools have been around for several years, but they often required coding knowledge, and the most efficient ones required significant computing power. This is now a thing of the past as the tools are very accessible.
And these AIs will soon be able to do more than create images, as every week sees great new possibilities, such as creating interior designs or 3D models – which can be used, for example, in the design of video games.
Even the web giants have joined in: Google and Facebook recently demonstrated systems (not yet publicly available) for creating entire videos from text, and TikTok already offers a text-to-picture tool.
And AIs don’t just do art and cinema, as some also create music or even entire podcasts.
But nothing guarantees that this crazy pace of progress will last forever: “We already had an ‘AI winter’ in the 1960s, with the sector alternating between phases of optimism and pessimism,” says Vincent Schmite.
It remains to be seen how most people will decide to get hold of it: “It’s not because of the democratization of smartphone camera use that everyone has become a professional photographer. Likewise, just because anyone can create an image from text doesn’t mean everyone will become an artist.”
Basic online creation tools
• Dall-E 2Created by the OpenAI Foundation
• Available via Midjourney Discord server
• Stable Diffusion, an open source software that can be used or downloaded for free through various sites such as DreamStudio or PlaygroundAI (but requires coding knowledge and a relatively powerful computer)