Tech

Inside Midjourney, The Generative Art AI That Rivals DALL-E

With a short text description, the tool can create surreal sci-fi environments and stylistic fine art—for a price.
Janus Rose
New York, US
Images generated by Midjourney showing yellow Minions against a pastoral background (Left) and an abandoned office building with large windows (Right)
Images generated by Midjourney

AI tools like DALL-E that generate images from short text descriptions are all the rage right now, so it only makes sense that other systems are starting to compete for space in the budding AI art market.

Last week, the company behind the generative AI art tool Midjourney opened its closed beta, allowing anyone to sign up for an account and generate surreal renderings—albeit with a limited number of “free” generations. The influx of new users was a stress test for the system, which rivals DALL-E in its ability to create stylistic and sometimes hauntingly realistic renderings. 

Advertisement

“For the next 24 hours we’re going to try and open up beta access to everyone,” wrote David Holz, Midjourney’s CEO, in a message posted to the AI tool’s Discord server on Monday night. “We don’t know if our servers are ready, but we’re ready to find out.” 

Unlike DALL-E, Midjourney seems to particularly excel at creating environments, especially fantasy and dystopian sci-fi scenes with dramatic lighting that look like rendered concept art from a video game. 

Image generated by Midjourney with the prompt: "Interior of an abandoned office building that has been used to throw secret raves"

Image generated by Midjourney with the prompt: "Interior of an abandoned office building that has been used to throw secret raves"

Images generated by Midjourney with the prompt: "interior of a spaceship filled with lush plants"

Images generated by Midjourney with the prompt: "interior of a spaceship filled with lush plants"

It’s also quite good at making surreal pastiche that mimics different art styles. Here are some examples where I tried to replicate classic art with celebrities like Megan Rapione and, uh, Minions:

Images generated by Midjourney with the prompt: "Megan Rapinoe piloting a Gundam in German Impressionist style"

Images generated by Midjourney with the prompt: "Megan Rapinoe piloting a Gundam in German Impressionist style"

Image generated by Midjourney with the prompt: "Photorealistic Minions inside the Windows XP background"

Image generated by Midjourney with the prompt: "Photorealistic Minions inside the Windows XP background"

The test gave way to a fully-open beta on Wednesday, allowing anyone to sign up and join the project’s Discord channel. The beta is operating entirely through Discord, with users typing their prompts directly into the chat interface and receiving messages from a bot that shows their generations rendering in real time. Users can then choose to upscale and enhance an image from each set of generations, or create more variations from the same prompt. 

However, this “free trial” period only gives each user a limited number of generations before the bot prompts them to buy a subscription. For non-commercial use, the cheapest plan allows 200 images for $10 per month, while the premium tier allows unlimited generations for $30 per month. (The makers of DALL-E also plan on offering paid access to their AI tool, and recently surveyed its beta testers on possible price points.)

Advertisement

Midjourney also—thankfully—discourages people from minting NFTs by requiring that anyone using generated images in “anything related to blockchain technologies” pay a 20 percent royalty on any revenue over $20,000 per month.

It’s still not clear how the hype around generative AI art will translate into a business model, however—especially when the large AI models they’re based on have been repeatedly shown to replicate racist and sexist stereotypes.

While the makers of DALL-E have attempted to mitigate some of the training biases inherent to these models, Midjourney hasn’t published any information about what datasets and methods were used to train its AI tool, and doesn’t seem to have many explicit content protections aside from automatically blocking certain keywords.

The “Content and Moderation” section of Midjourny’s user guide instructs users to “not create images or use text prompts that are inherently disrespectful, aggressive, or otherwise abusive,” and to “avoid making visually shocking or disturbing content” including adult content and gore. The rules also forbid content that “can be viewed as racist, homophobic, disturbing, or in some way derogatory to a community,” including “offensive images of celebrities or public figures.”

It’s unclear how or how well any of this will be enforced, but with such impressive results, the Midjourney project will likely be something to watch as companies start precariously navigating a path forward for image-generating AI.