They create a spectacular photo or a sophisticated drawing in a few seconds without the need for human genius: image generators with AI require neither artistic nor creative skills, except that the user is imaginative in his commands to the machine. You simply enter instructions as text: "Spiderman, Donald Duck, Joker, Popeye, Superman and Wonder Woman greet the Pope in front of a backdrop of Catholic nuns. There is a church in the background." A few minutes later, the software has created four possible motifs and displays them in a preview image. You select one, which is now displayed in high resolution. The result can be seen at the top left. Wonder Woman, Donald Duck and the nuns are missing, and the algorithm probably couldn't do anything with Popeye.

Michael Spehr

Editor in the "Technology and Motor" department.

  • Follow I follow

The photo shows the strength and weakness of the image generators. Like chat systems with artificial intelligence, such as ChatGPT or the new Microsoft Bing, the software is based on continuous training with material from the Internet. In this case, millions of images. The AI "learns" from the image descriptions what a dragon or a dog is, what Superman looks like and a church. Learning takes place with neural networks and, as with ChatGPT and Microsoft Bing, the same applies here: The results cannot be relied upon, because the AI machines do not "know" anything, but work with statistical relationships.

What you enter as a command is at best generated that way

The photos shown here were taken with the software Midjourney of the American research institute of the same name. The app can be used in beta software on a Windows or Mac computer, and its commissioning is not easy. You need an account of the online service Discord and install Midjourney there as a bot. As with its siblings Dall-E or Stable Diffusion, Midjourney is programmed at a command prompt with text commands. What you enter as a command is at best generated that way. You can enter long, detailed phrases and specify what you do not want to see in the picture. Anything that could become obscene is immediately rejected by the machine.

Also, public figures cannot be integrated into the artificially generated works of art. We tried it anyway and pretended: "Robert Habeck and Volker Wissing go on holiday by the sea as best friends." The two bearded men on a boat with seagulls flying over is the result. We assume that the AI could not do anything with both names.

Also, the command "The most popular German politicians stand by a swimming pool and smile into the camera" was not carried out correctly. The software drew well-dressed people standing dressed in a pool.

In addition to the description, an art style and a mode of representation can be specified. On the left, for example, we gave a "Horses fortify themselves at a gas station in the style of Edward Hopper". How well the painter Edward Hopper was met here remains to be seen. The "robot dogs in a meadow, as Hieronymus Bosch would have painted" also pushed the AI to their limits.

Midjourney allows you to artistically alienate your own photos. You upload some of them, a URL is generated from the image data set, which in turn can be embedded in scenes. Because the creation of photos and graphics requires a lot of computing power, you sometimes have to wait for minutes for a first result. The software can be tried out free of charge, for longer use a subscription is indispensable. Midjourney can also be tried out on the iPhone with the app "AI Art" (for 23 euros per month) in a limited range of functions.