T
he whole Internet speaks about Midjourney the art that this neural networkcreates is easily confused with the works of professional artists. At the same time, the work of artificial intelligence is thousands of times cheaper, and the time to create paintings is measured in seconds.
How does this diabolical machine work from the inside, who invented it and why does the world need living designers and artists if there is a Midjourney?
Scientist, startup, revolutionary
Midjourney is a neural network developed by the American company of the same name, which amazed users around the world with pictures (otherwise it can not be called) created on the basis of text queries. In February 2022, the project was founded by scientist and entrepreneur David Holtz, a 33—year-old graduate of the Faculty of Applied Mathematics at the University of North Carolina at Chapel Hill. As a student, Holtz managed to work at the Max Planck Institute, where he studied neuroimaging algorithms and developed a map of the rat brain at the cellular level, and even at the NASA research center, where he was engaged in LiDAR technology.
In 2011, the young scientist left graduate school and moved to San Francisco, where he founded his first company, Leap Motion (now Ultraleap), which develops motion sensors and other human gesture recognition systems. Holtz ran the firm until 2021, but decided to go out of business: he, according to his own words, did not want to run a large company — in an interview with The Verge, Holtz admitted that he was interested in another, young and rapidly developing environment.
In early 2022, Holtz resigned from the Ultralap founders and founded Midjourney. According to the entrepreneur, the staff of the company even now, when it has overtaken world fame, does not exceed 10 people, the project has no investors, and money is not the main motivation of the founder. "The main thing I want is that for the next 10 years we have a home where we can experiment with technologies and create products that will matter not only to me, but to the whole world. Well, have fun in the process," he says.
How it works?
The work of Midjourney is provided by two technological breakthroughs in the field of artificial intelligence that have occurred relatively recently: the ability of neural networks to understand human speech and create images. In order to transform these two skills into a coherent system that produces works of art on request, the neural network is trained to build a correspondence between textual descriptions and visual images on hundreds of millions of examples. The results of such training allow us to solve various cross-modal tasks — the generation of images by text description, the generation of text descriptions by pictures, the completion of parts of the image, and so on, says Sergey Markov, head of the department of experimental machine learning systems at SberDevices. "Midjourney is a diffusion neural network and consists, as it were, of two neural networks: the first is responsible for processing and understanding text, the second for generating images," explains Markov.
In mid-July, Midjourney entered the beta testing phase and became available to users around the world. However, to give a task to Midjourney, you need to be registered in Discord, a cross—platform messenger popular with gamers, game developers and designers. First you need to go to the official Midjourney website and log in via Discord, then pay for a subscription or use the free version. The free version allows you to generate and download 12 images, but does not give access to your personal account (this prevents you from tracking the fate of your requests in the general chat), for $ 10 you can create up to 200 images per month, for $ 30 you can generate an infinite number of images. A $600 corporate subscription is also available, which gives company employees the opportunity to create pictures in a team and view each other's individual work.
According to Holtz, he chose the way to access the system via Discord because of the group principle of the platform: people are more willing to fantasize when they gather in groups, Holtz believes. By joining the service, you can send text commands to create images together with other users or singly on any of the many Discord channels.
To create an image, it is enough to enter into the chat with the Midjourney bot words describing the picture that you want to get in the end. The system will generate four images to choose from, and then the most suitable image can be scaled, modified and refined to the ideal.
The resulting images appear in the general Discord channel about a minute after the request is sent. Holders of a paid subscription can send commands to the bot in the format of private messages, and not through a public channel. But the images generated by the neural network remain publicly available for viewing by default.