How Midjourney works, the AI ​​that creates images from your descriptions

Non si possono proteggere con il copyright le immagini AI di Midjourney thumbnail

From the word to the images, from the idea to its realisation. All this using artificial intelligence. This is the promise of Midjourney, text-to-image artificial intelligence capable of generating images starting from text: what is this tool and how does it work? And how to use it to create images with AI?

What Midjourney is and how it works, the AI ​​that generates images

Midjourney it’s a’text-to-image artificial intelligencei.e. a system that uses deep learning algorithms to create realistic imagesand starting from a text description. Basically, just write what you want to see and Midjourney generates it in no time. You can choose the level of detail, color, style and composition of the image. You can also combine different elements or modify existing ones. It’s like having a graphics expert at your disposal: even if the result will not always be as precise, and almost never as well studied.

Midjourney was created by an independent research laboratory based in San Francisco, founded by David Holzwho had also created the company Leap Motion (specialized in sensors and trackers for virtual reality).

Midjourney is available from March 2022 through social media Discord. Today there is also a site through which to access the bot. Be careful though: after using the first 25 minutes of GPU (the “unit of measurement” for image generation, one minute is roughly equivalent to one request), you have to sign up for a subscription (Standard plan is $10/month, Pro is $60/month).

How Midjourney works

Midjourney one first artificial neural network called Generative Adversarial Network (GAN)which consists of two modules: a generator it’s a discriminator. The generator has the task of producing the images starting from the text. While the discriminator has the task of evaluating whether the images are realistic or not. The generator tries to fool the discriminator, while the discriminator tries to recognize the fake images. In this way, the two modules train each other and improve their performance.

Midjourney is based on an advanced version of GAN called CLIP+DALL-E+VQGAN, which combines three components: CLIP (Contrastive Language-Image Pre-training), GIVE HER (Deep Autoregressive Language Model for Image Generation) e National Park (Vector Quantized Generative Adversarial Network). CLIP is a neural network that associates words with images based on their semantic similarity. DALL-E is a neural network that generates images from words using an autoregressive model. VQGAN is a neural network that compresses images into discrete vectors and improves their quality.

What to use the AI ​​image generator for?

how does midjourney min

Midjourney is a very powerful and versatile tool for text-to-image image generation. Can be used for purposes artistic, educational, recreational or informative. He can create never-before-seen scenes, combine elements in original ways, or modify existing ones. It can also produce hyper-realistic images that look like real photographs. This in particular has led to the spread of various fake news, such as those recently seen with Pope Francis’ jacket or the fake arrest of Donald Trump.

However, Midjourney also has limitations and risks. First, you are not always able to correctly interpret the text or meet the specifications requested by the user. TO sometimes it generates wrong, incomplete or inconsistent images. We therefore need attention on the part of users to use it to the fullest.

How does Midjourney work on Discord?

Discord is a communication application that allows you to exchange messages, calls and video calls with other people. She was born in 2015 and had very successful among gamers, who use it to coordinate and socialize. Among its features there is also the possibility of create bots, i.e. programs that interact with users via written commands.

One of these bots is Midjourney, which uses artificial intelligence to generate images from a text description. To use Midjourney, you must first sign up for Discord and then enter its server, where there are several dedicated channels. To access the server you can use this invitation link, which leads to a community of over 13 million users interested in this technology. Click “Accept invitation”.

Using Midjourney takes several tries

However, using Midjourney is not as simple as it seems. You often have to try several times to get the desired image, by modifying the description or the parameters. Many artists or creative industry professionals use Midjourney as source of inspiration or as a tool to explore different options.

midjourney how it works min

As reported by the Corriere della Sera, Jos Averya photographer who posted AI portraits on Instagramto make 200 photos, generated more than 13,000 images with Midjourney.

To try Midjourney you need to enter one of the channels called “Newbies“, which have an associated number. There you can write your own description in the space provided eaWait 60 seconds to see four possible outcomes.

How to write the prompt

To use Midjourney, you have to write your description after the command “/imagine” (without the quotes). The description must necessarily be made in English. Don’t be afraid to add details, rather write longer, to give more material to the AI. But be prepared, to get what you want you may have to repeat the process several times.

Midjourney then generates four images, all quite similar to each other, but with some different details. You can enlarge the images by clicking the button (U), what does it mean “Upscaling“: remember that the numbers go clockwise starting from the top left.

You can also click (V) to see a new version of a particular image, or (Refresh) to have four new images. By clicking on Open in browseryou can save the image like any other photo on the web.

The extra paid commands

There are several extra possibilities for those who use one of the paid versions. For example, you can use the commands:

/fast = to create an image faster
/private = not to show our image to other users
​-he -w = to give the precise dimensions of the image
-ar = to choose the format
-video = to ask to create a video
-quality 2 o -quality 5 = to increase the image quality
-stylize625 or stylize1250 (and then again 2500, 20000 and 60000) = to decide the artistic level of the image

With these extra options you have more control, but some require an extra payment on top of your subscription.

Some rules and limits

Given the use of Midjourney images to spread fake news, the development team has introduced some limits. For example, after the fake arrest of Donald Trump you can’t ask for pictures of famous people in handcuffs or in prison. But you can still ask for clashes between famous personalities and many other potentially controversial images. But expect new rules and limits in the future: the company that operates Midjourney is still figuring out how to adjust the policies.

You should now have a clearer idea of ​​how Midjourney works and how to leverage it on Discord. But as we said, it will take practice – and many wrong images – to be able to take full advantage of it.


Corriere della Sera