MidJourney is an artificial intelligence tool that generates images from texts. However, unlike Dall-E, this AI has yet to be trained to aim for realism but rather for artistic beauty. The result is stunning, and many artists are impressed. Find out everything you need to know.
Since June 2022, the artificial intelligence program Dall-E Mini alias Craiyon has met with viral success on social networks. This AI can create images from user-entered text and has spawned many hilarious memes.
A few weeks after Dall-E Mini, a new “text-to-image” generator based on artificial intelligence is creating a buzz on the internet. This is MidJourney.
What is MidJourney?
MidJourney is an artificial intelligence system capable of creating images from the text entered by the user. On the official website, its creators present themselves as “an independent research laboratory. Exploring new ways of thinking. Extending the powers of the imagination of the human species“.
This AI model was trained on many images, like most text-to-image AIs. However, it is distinguished by its emphasis on artistic style rather than realism. By optimizing their AI, the creators of MidJourney wanted, above all, for the images generated to look good.
In version 3.0, the software currently incorporates a feedback loop based on user activity and reactions. It has improved image quality by analyzing data indicating which images users like and how they use them.
The company has yet to reveal its technologies but confirms that it uses the most advanced artificial intelligence models with billions of parameters and is trained on billions of images. Also, the images are generated on a vendor’s cloud using green energy.
Each image requires several pet apps, which is 10^15 operations per second. According to the software creators, no service accessible to the general public has ever allowed an individual to use so much computing power.
What is MidJourney used for?
Most users use this tool for fun and to bring their imagination to life. However, about 30% of users have professional use of this tool.
Many graphic designers use MidJourney as part of their concept development workflow. They generate several variations of an idea and present it to their clients to determine which direction to pursue. According to Holz, professionals use this tool to “supercharge” the creative or communication process.
In addition, still according to the creator, about 20% of users would use it for therapeutic purposes. Creating images could help them overcome trauma or bereavement. AI is a “tool for emotional and intellectual reflection” here.
For a long time, MidJourney was for non-commercial use only. From now on, commercial use can be authorized. In June 2022, the famous magazine The Economist let AI create its cover.
How to create images with MidJourney?
Currently, MidJourney is still in beta. To use it, you need to register on the official website. Invitations are then sent out in waves.
After receiving an invitation email, you will communicate with the program via Discord messaging on a public channel. Enter some text, and the MidJourney robot will return an image after about a minute.
However, paying a subscription for 10 dollars or 30 dollars per month to submit your text in private via a direct message to the robot is possible. It avoids messages from other users of the public channel. By default, images generated by the AI nevertheless remain publicly visible.
What content is prohibited?
As a social app, MidJourney sets rules and limits regarding allowed content. Gore or adult content is strictly prohibited, and users are urged to avoid creating visually offensive or disturbing content.
Specific texts can thus be blocked automatically. In addition, about forty moderators keep an eye on the images users create.
The creators of MidJourney also don’t like the idea of their AI being used to create fake DeepFake-like photographs. According to them, this is a “hazardous” use.
Who owns the intellectual property of the art created by MidJourney?
The issue of intellectual property is sensitive when discussing art generated by AI. Currently, American justice prohibits, for example, granting copyrights to AI-generated images.
In February 2022, the U.S. Copyright Office Review Board denied a motion to copyright a computer-generated landscape image titled “A Recent Entrance to Paradise .” For a good reason, a human had not created this image.
According to MidJourney’s terms of service, users own all works it creates with the service. However, the company requires a copyright license from users to reproduce content created with the service. It is a necessary precaution for hosting user images.
Additionally, intellectual property issues could arise regarding AI models trained on copyrighted material. An AI trained on existing content also risks generating images with similarities…
The creators of MidJourney recognize that the law still needs to be more precise regarding AI-generated content. He expects more specific laws to be passed in the future.
David Holz: who is the creator of MidJourney?
David Holz grew up in Florida. He started his design business while studying math and physics at university. He was preparing for his Ph.D. in Mathematics but finally took a break in 2008 to co-found Leap Motion. The latter was dedicated to the manufacture of computer peripherals.
Holz spent a year as a research student at the Max Planck Institute the following year. He then spent two years as a student researcher at NASA’s Langley Research Center. He has also worked on LiDAR, Mars missions, and atmospheric science.
However, this time was also a time of doubt for Holz. He explains that he wondered “why he was working on these things.” After that, he wanted to work on something close to his heart.
So he focused on Leap Motion. The latter develops a hardware device allowing one to follow the hand’s movements and use them to control a computer or any other device.
Holz ran this business for twelve years. When he left, it employed about 100 people. In 2021, he finally packed up to create MidJourney.
He says this company is still relatively small and has only about ten people. It is also self-funded and has no investors.
Holz explains that the team’s motivations could be more financial. He said, “we’re just here working on things we’re passionate about and having fun. And we’re working on a lot of different projects .”
According to him, the technological aspect of AI and the extent to which it will improve is relatively easy to predict. On the other hand, “the human ramifications of this extension are very difficult to imagine. “
In his eyes, “there is something here at the intersection of humanity and technology. To really understand what it is and what it should be, we really need to conduct a lot of experiments“. Holz, however, points out that the company is already profitable.
MidJourney vs. DALL-E
While Dall-E focuses on the images’ realism generated, MidJourney emphasizes the artistic dimension more. This AI can adopt real art styles to create an image from a combination of any desired elements.
To demonstrate the difference between the two, a user named Fabian Stelzer typed in the exact text on both programs.
He says, “MidJourney has a certain ‘je ne sais quoi’, the imperfections are more beautiful, a bit like on an analog synthesizer. It often turns out to be more contextually creative and great with textures. DALL-E is best for scenes with obvious instructions.”
Texts submitted to the two IAs for this test included “Mozart playing Top of the Pops, 1993” and “a complex installation made of plastic bags and mirror fragments, painted in neon color, studio lighting. »
MidJourney succeeded in creating a much more realistic image for the installation of plastic bags. The lighting of the studio is particularly successful.
The phenomenon of AI-generated art
This fashion for image-creating AIs began in 2021 with the launch of CLIP (Contrastive Language Image Pre-Training) by OpenAI. This program was designed to assess how well-generated images line up with textual descriptions, but then artist Ryan Murdock realized the process could be reversed: an AI can produce an image from text.
The generative art community then began a period of unbridled exploration, releasing Python code for creating images using a variety of patterns and techniques.
According to David Holz, creator of Midjourney, it was around this time that “some areas of AI were progressing in exciting ways, and one of them was the ability of AI to understand the language.
It refers, in particular, to transformers: a type of Deep Learning model based on CLIP. Similarly, diffusion models are an alternative to GANs (antagonistic neural networks).
This expert was particularly impressed by the diffusion guided by CLIP, developed by Katherine Crawson, known on Twitter under the pseudonym @RiversHaveWings.
The future of MidJourney
For the future, the creators of MidJourney fear it will be a victim of its success. Currently, several hundred thousand people are already using the service and require the power of around 10,000 servers.
However, if 10 million people attempt to use this technology, Holz explains, there won’t be enough computers worldwide.
According to Holz, AI tools like MidJourney can help artists improve their discipline. They are not necessarily intended to make everyone a professional artist, and an artist using these tools will always be better than a simple individual using them.
However, he acknowledges that these tools could add pressure on artists, even if this is different. He believes that the tool will improve enormously over the next two years.
MidJourney’s best creations
By creating a MidJourney account, you can view other users’ creations on the Community Feed page. New AI-generated images are constantly pouring in. Here is a selection of impressive or surprising creations made by MidJourney and relayed on social networks.
MidJourney V4: all the new features
Since Saturday, November 5, 2022, MidJourney has started the alpha test of its V4. This new version offers more detailed results and makes obtaining high-quality results from simple texts easy.
This major update succeeds the V3 available since August. Several thousand members of the official Discord server can test this fourth version simply by adding “-v 4” to their prompts.
According to founder David Holz, “V4 is an entirely new code base and an entirely new AI architecture. This is our first model trained on a new AI Midjourney supercluster that we have been working on for 9 months“.
The images produced by this V4 model are much more detailed. The prompts seem better built, the composition of the scenes improved, and the proportionality also seems more realistic.
Another novelty is a significantly increased knowledge of places, creators, and more. Small details are much better represented, and complex promptings with multiple levels of detail are better supported.
This V4 is more suitable for scenes with multiple objects or people and offers advanced features such as multi-prompt or image prompting.
Generally speaking, this update is warmly welcomed by users. It is an alpha, and the new model will continue to be improved over the coming weeks. The firm plans to increase the definition and quality of upscaled images, add custom aspect ratios as on V3, increase the image’s precision, and reduce text artifacts.
MidJourney bans words related to the reproductive system to avoid porn content.
The generation of pornographic content has always been prohibited on MidJourney. However, users are savvy and have quickly found techniques to circumvent this restriction.
At the end of February 2023, the platform finally decided to ban words related to the human reproductive system from fighting against these abuses.
From now on, terms such as “placenta,” “fallopian tubes,” “mammary glands,” “sperm,” “uterine,” “cervix,” “hymen,” and “vulva” can no longer be used as prompts. Sometimes, the user attempting to use one of these words will be blocked for a limited time.
According to David Holz, this is a temporary measure to prevent the creation of violent or sexualized images while the company improves AI. The list of forbidden words can be consulted on the page dedicated to MidJourney’s terms of use.
Cyberpunk: Peach John, the first manga created with MidJourney
The manga “Cyberpunk: Peach John” will be released in Japan. However, its author, nicknamed Rootport, admits to having no talent for drawing.
To create his comic, he turned to MidJourney and other AI tools like Stable Diffusion and DALL-E 2. In just six weeks, he created a manga of over 100 pages.
Within a minute, Rootport had already created the hero of his story using a prompt containing keywords like “pink hair,” “Asian boy,” or “stadium jacket.”
He then assembled the best images into a comic book format to produce a book. On March 9, 2023, this work will be published by the famous publisher Shinchosha.
According to the author, generative AI could allow many people without the artistic talent to enter the manga industry if they have good stories to tell.
According to him, this creative process is akin to a lottery or the enchantment of magic spells allowing him to generate images from his imagination. He admits, however, that the satisfaction of drawing oneself is undoubtedly greater.
Of course, this first AI-drawn manga raises concerns for this multi-billion-dollar industry. Thus, Professor Satoshi Kurihara confides to AFP that the junior assistants of mangakas risk being replaced by AI …
MidJourney goes to version 5: photorealism and (almost) regular human hands.
On Wednesday, March 15, 2023, MidJourney announced the launch of version 5 of its text-to-image service. This new version can produce photorealistic images, so much so that some users find them almost “too perfect.”
There is a clear improvement compared to the first version launched in March 2022, but also to version 3, launched in August, and version 4 in November. Each iteration adds more detail to the generated results.
Skin textures and facial features are very realistic, as is the lighting. Highlights and shadows are better, and viewing angles are more successful.
Also, the eyes are now almost perfect, and the hands are finally (almost) normal. It is a big step forward, as this was one of the main weak points of the tool so far.
In addition, MidJourney specifies that v5 offers a much more comprehensive range of styles and is more sensitive to prompting. It also generates less unwanted text and offers twice the image definition.
MidJourney v5 is available in alpha testing for subscribers to the service. It can be accessed from Discord.
MidJourney Launches Entirely AI-Created Magazine
Following consultation with its community, MidJourney launched an eponymous monthly magazine in paper format.
Each month, readers will be able to discover a selection of images created using AI, highly rated by users. The post will also contain sample prompts and interviews with the image creators.
The subscription will be charged 4 dollars per month, but the first subscribers will be able to receive the first issue for free.