Artificial intelligence, which generates images based on text requests, became a trend among Internet users this year. Among the most popular is Midjourney, a neural network that works as a chatbot in Discord. It was created in February 2022, but has already managed to go viral and turn into an exciting toy. This is not the first and not the only service for generating pictures based on text requests, but it is so simple that any smartphone user can "draw" with it, without even needing a computer. Perhaps this is the secret of his success.
We've studied and tested this fancy AI in detail to tell you how to use Midjourney, how to manage requests, and what fun things you can do to co-author it.
The neural network draws like a real artist
Go to the Midjourney website and select the Join the beta section.
The system will transfer you to Discord and ask you to join the neural network channel. You need to have an account on this service or pre-register for it.
Go to any newbies chat and type /imagine prompt: in the message field, or just type /(slash) and select the imagine command from the list.
Write a request in English in the field. Wait until the system generates 4 variants of the picture.
Choose one of them to make a few more variations or immediately detail as an independent image.
Download the image to the gallery right away, because many other users are working with you in the same chat, and your work will "go" far up in a few minutes.
You have 25 free actions, which include image generation, variant creation and final processing. Use them.
Mastering Midjourney - how to use a chatbot in Discord
If you want to continue using the service, you can sign up for a paid plan.
To do this, you need to enter the /subscribe command in the message field and the bot will transfer you to the subscription page. There are two basic plans that are sufficient for normal users:
Basic Membership (Basic Membership) - costs $10 per month, provides 200 minutes of fast image generation (gpu-minutes).
Standard Membership (Standard Membership) - costs $30 per month, provides an unlimited plan for rapid image generation. In fact, it means 15 gpu-hours, which is a lot. But even if you use them, on this tariff there is an opportunity to work in a relaxed mode, for which paid minutes are not deducted.
Fast generation mode --fast means that the system gives you priority in the queue for using service resources.
Relaxed mode --relax means that your images will be generated longer, because you will queue up for other users. However, it also happens quite quickly. This option is only available in Standard and Corporate plans.
The developers note that if you lose your mind and generate too much, you will automatically be put in slow mode and put at the end of the queue.
You can find out how many minutes you have left on the account and other information about the account using the /info command.
After the system has generated 4 image thumbnails, you can choose two actions - V (make more options) or U (enlarge and detail, this button makes the final version of the image). For example, if you press V3, the system will make 4 more versions of the bottom left picture.
Hint: if you use the V command several times, the AI will produce something very strange. have fun
The neural network draws online in real time - you can see how the picture is created at your request
When you get the final image after pressing U, you can continue to work with it further by using one of the buttons - Make Variations allows you to make 2 more versions, Upscale increases the quality, Remaster makes the image in a different style.
The finished image is uploaded to a dedicated neural network server and is available via a link that is immediately displayed in the chat along with the generated image.
Working with artificial intelligence can quickly get boring if you don't know how to write queries correctly. In addition to the usual text description of the desired picture, such as "cat, red-haired, on the sofa", there are a number of other parameters that can be used to specify the image format.
Give the AI a drawing style - in the style of Salvador Dali, Hieronymus Bosch, Da Vinci, cyberpunk, pop art, realistic, detailed.
A cat drinks milk in the style of Salvador Dali
Adjust lighting and colors using words: natural light, soft light, diffuse, sharp shadows, etc.
Set the format and size of pictures. The --ar prompt allows you to set the aspect ratio. By default, the service generates square images, but you can make horizontal or vertical images by setting parameters, for example, --ar 3:2. You can specify a specific width or height in pixels using the --w <number> (width) and --h <number> (height) hint.
Choose image quality (detail). The --quality <number> or --q <number> prompts instruct the system to decrease or increase image quality. The default value is 1, but you can choose 0.5 to decrease or 2 to increase. If you choose a value of 5, you can get something weird.
Experiment. With the help of commands --chaos <number from 1 to 100> and --stylize <number from 625 to 60,000> you can make versions more diverse, creative and unexpected.
You can learn more in the section on the site, where the main commands for the bot are given. All additional operators with two dashes (--) must be placed at the end of the query.
In addition, it is worth formulating the description of the image itself with a creative approach. Artificial intelligence sometimes understands the task in a unique way and can produce something completely unexpected. In fact, this is a game. Therefore, you will get the most interesting result by giving an unusual request:
Lines from poems or songs;
Abstract concepts (mutual love, happy childhood, love for the native land, punishment for wrongdoing, etc.);
Oxymorons are phrases that simultaneously contain two mutually exclusive concepts (hot ice, sad fun).
This is how artificial intelligence saw the phrase "hot ice".
The more creative and strange the query you create, the more original and interesting the result you can get.
Of course, we were interested in giving the neural network queries that are relevant to our field of activity. We chose the most abstract terms that are difficult to visualize. This is how artificial intelligence saw the virtual server .
So Majorni saw a virtual server
We also asked the AI to render the domain . The first attempt was not very successful - the bot gives out fortresses and cities. But this is not surprising, because the word domain in English also means "possession", "plot", "property". Therefore, the task was complicated and a request for domain name registration was given. And here the neural network surprised. What is MAM and where does such an association come from is a mystery.
Midjourney is a community where images are generated in public chat (unless you've paid for private use and no one but you can see them). The developers feel a moral responsibility for the generated content, so they have created a number of prohibitions.
Yes, the bot has a list of words that are automatically blocked if they refer to pornography, bloody and violent scenes, as well as anything that can offend other participants, cause them fear or disgust.
When you enter such a word in the query, the bot gives a response about blocking the command. If the user often accesses prohibited keys, he can be completely blocked.
Bans apply to all users, even those who have signed up for a private subscription.
The copyright situation in Midjourney is twofold. On the one hand, the copyright for the generated content belongs to the user who created the picture. On the other hand, the service retains exclusive rights to use, edit and display all materials that pass through the chat. Each member of the chat can see your requests and results, and can also mix and process them. That is, all the buttons under your generation, described above, can be clicked not only by you, but also by anyone in the chat.
This means that no one will sue you for, for example, selling posters with images generated in the bot. And on the other hand, the same pictures can roam freely on the Internet.
A neural network that generates images is a great toy, but is it of any practical use? Of course, it can also be used for work:
Ideas for artists, illustrators, animators;
Illustrations for books and individual works;
Illustrations for music albums;
Ideas for worlds and characters (games, movies, novels);
Creation of posters and printed interior pictures.
In general, a neural network produces fairly similar results if you work with it for a long time, and they need to be refined and developed. Therefore, the question of whether the neural network will replace illustrators remains open. Artificial intelligence does not currently have enough imagination to compete with a human, but at the same time it is crazy enough to produce something unexpected. However, artificial intelligence is developing, so time will tell.
Although the neural network is just a service that does not have a soul, its own thought or character, we can conditionally talk about some of its "chips", which are traced in all the pictures.
Absence of faces, people with their backs turned
The neural network tries to avoid drawing faces if the request does not provide for it. For example, if you query for "passengers," "people," or "crowd," you'll get all the options without faces. Some requests may contain versions with and without faces (warrior, childhood, god). Faceless images may be easier to generate, but you might think that AI is a misanthrope.
In almost all images where faces are present, you will see strange, deformed eyes that are difficult to describe. They are so specific that you can guess the "authorship" of the image by looking at it - it was drawn by artificial intelligence.
Distorted and twisted forms
For humans, there are basic parameters of living and non-living world objects, and it is difficult for us to go beyond the usual limits. The neural network does not have such parameters, for it, surrealism is a way of thinking. Therefore, it can generate planets of any shape except round, a bird with one wing, or a woman with arms of different lengths. Despite the fact that there was nothing similar in the request. This is explained by the fact that the neural network is not yet sufficiently trained.
If you do not specify colors in the request or there are no objects to which certain associations are attached (winter is white, blood is red), then the neural network itself chooses the colors. Most often they are gray or turquoise-red.
Request "Mutual Love"
By working with an AI for a while, you can learn more about its "character" and use that information to improve your drawings.
We haven't had time to marvel at the images yet, and Meta has already announced that its neural network has learned to generate video on demand. The service is not yet available for wide use, but the creators showed several short videos.
They are as "stubborn" as anything artificial intelligence does.
It is planned that with their help it will be possible to edit videos, animate photos and do many useful things.
Neural networks are booming, weird and imperfect, but extremely exciting. The fact that ordinary Internet users were able to join their use indicates a new turn in the history of artificial intelligence. Let's see what happens next.