An interview with David Holz, CEO of AI image-generator Midjourney: it’s ‘an engine for the imagination’

‘An engine for the imagination’: the rise of AI image generators

image

An AI-generated image was made with the following prompt: “A robot steam engine, barreling down the tracks at terrific speed, emitting clouds of colorful smoke.”

AI-generated artwork is quietly beginning to reshape culture. Over the last few years, the ability of machine learning systems to generate imagery from text prompts has increased dramatically in quality, accuracy, and expression. Now, these tools are moving out of research labs and into the hands of everyday users, where they’re creating new visual languages of expression and — most likely — new types of trouble.

There are only thought to be a few dozen top-flight image-generating AI in existence right now. They’re tricky and expensive to create, requiring access to millions of images used to train the system (it looks for patterns in the pictures and copies them) and a great deal of computational grunt (for which costs vary, but a million-dollar price tag isn’t out of the question).

Right now, the output of these systems is mostly treated as novelty when it gets splashed on a magazine cover or used to generate memes. But as we speak, artists and designers are integrating this software into their workflow, and in a short amount of time, AI-generated and AI-augmented art will be everywhere. Questions about copyright (who owns the image? Who made it?) and about potential dangers (like biased output or AI-generated misinformation) will have to be dealt with quickly.

As the technology goes mainstream, though, one company will be able to take some credit for its ascendancy: a 10-person research lab named Midjourney, which makes an eponymous AI image generator accessed through a Discord chat server. Although the name might be unfamiliar, you’ve probably seen the output from Midjourney’s system floating about your social media feeds already. To generate your own, you just join Midjourney’s Discord, type a prompt, and the system makes an image for you.

“A lot of people ask us, why don’t you just make an iOS app that makes you a picture?” Midjourney’s founder, David Holz, told The Verge in an interview. “But people want to make things together, and if you do that on iOS, you have to make your own social network. And that’s pretty hard. So if you want your own social experience, Discord is really great.”

Sign up for a free account, and you get 25 credits, with all images generated in public chatrooms. After that, you’ll have to pay — either $10 or $30 a month, depending on the number of images you want to make and whether or not they’re private to you.

This week, though, Midjourney is expanding access to its model, allowing anyone to create their own Discord server with their own AI image generator. “We’re going from a Midjourney universe to a Midjourney multiverse,” as Holz puts it. And he thinks the results will be incredible: an outpouring of AI-augmented creativity that’s still only the tip of the iceberg.

To find out more about Holz’s ambitions with Midjourney — about why he’s building an “engine for the imagination” and why he thinks AI is more like water than a tiger — we rang him up for an interview. And, of course, we got Midjourney to illustrate our conversation.

This interview below has been condensed and lightly edited for clarity.

It’d be great to start with a bit about yourself and Midjourney. What’s your background? How did you get in this scene? And what is Midjourney — a company, a community? How would you describe it?

So, my name is David Holz, and I guess I’m a serial entrepreneur. My brief history would be: I had a design business in high school. I went to college for physics in maths. I was working on a PhD in fluid mechanics while working at NASA and Max Planck. I got overwhelmed at one point and put all those things aside. So I moved to San Francisco and started a technology company called Leap Motion around 2011. And we sold these hardware devices that would do motion capture on your hands, kind of inventing a lot of the gestural interface space.

I founded Leap Motion and ran that for 12 years, [but] eventually, I was looking for a different environment instead of a big venture-backed company, and I left to start Midjourney. Right now, it’s pretty small — we’re like 10 people, we have no investors, and we’re not really financially motivated. We’re not under pressure to sell something or be a public company. It’s just about having a home for the next 10 years to work on cool projects that matter —hopefully not just to me but to the world — and to have fun.

We’re working on a lot of different projects. It’s going to be a wide and diverse research lab. But there are themes: things like reflection, imagination, and coordination. And what we’re starting to become well known for is this image creation stuff. And we don’t think it’s really about art or making deepfakes, but — how do we expand the imaginative powers of the human species? And what does that mean? What does it mean when computers are better at visual imagination than 99 percent of humans? That doesn’t mean we will stop imagining. Cars are faster than humans, but that doesn’t mean we stopped walking. When we’re moving huge amounts of stuff over huge distances, we need engines, whether that’s airplanes or boats or cars. And we see this technology as an engine for the imagination. So it’s a very positive and humanistic thing.

Lots of labs and companies are working on similar technologies that turn text into imagery. Google has Imagen OpenAI has DALL-E and there are a handful of smaller projects like Craiyon. Where did this tech come from, where do you see it going in the future, and how does Midjourney’s vision differ from others in this space?

“in 10 years, you’ll be able to buy an Xbox with a giant AI processor, and all the games are dreams.”

So, there have been two breakthroughs [in AI that led to image generation tools]. One is understanding language, and the other is the ability to create images. And when you combine those things, you can create images through the understanding of language. We saw those technologies coming up, and we saw the trends — that these will be better at making images than people — and it’ll be really fast. Within the next year or two, you’ll be able to make content in real time: 30 frames a second, high resolution. It’ll be expensive, but it’ll be possible. Then, in 10 years, you’ll be able to buy an Xbox with a giant AI processor, and all the games are dreams.

From a raw technology standpoint, those are just kind of facts, and there’s no way to get around that. But from a human standpoint, what the hell does that mean? “All the games are dreams, and everything is malleable, and we’re going to have AR headsets” — what the hell does that mean? So the humanistic element of that is kind of unfathomable. And the software required to actually make that a thing that we can wield, it’s completely off the map, and I think that’s our focus.

We started off testing the raw technology in September last year, and we were immediately finding really different things. We found very quickly that most people don’t know what they want. You say: “Here’s a machine you can imagine anything with it — what do you want?” And they go: “dog.” And you go “really?” and they go “pink dog.” So you give them a picture of a dog, and they go “okay” and then go do something else.

Whereas if you put them in a group, they’ll go “dog” and someone else will go “space dog” and someone else will go “Aztec space dog,” and then all of a sudden, people understand the possibilities, and you’re creating this augmented imagination — an environment where people can learn and play with this new capacity. So we found that people really like imagining together, and so we made [Midjourney] social. And we have this giant Discord community, like it’s one of the largest Discords, with roughly a million people where they’re co-imagining things in these shared spaces.

Do you see this human collective as parallel to the machine collective? As a sort of counterbalance to these AI systems?

Well, there isn’t really a machine collective. Every time you ask the AI to make a picture, it doesn’t really remember or know anything else it’s ever made. It has no will, it has no goals, it has no intention, no storytelling ability. All the ego and will and stories — that’s us. It’s just like an engine. An engine has nowhere to go, but people have places to go. It’s kind of like a hive mind of people, super-powered with technology.

“They’re new, interesting, human aesthetics that I think will spill out into the world”

Inside the community, you have a million people making images, and they’re all riffing off each other, and by default, everybody can see everybody else’s images. You have to pay extra to pull out the community — and usually, if you do that, it means you’re some type of commercial user. So everyone’s ripping off each other, and there’s all these new aesthetics. It’s almost like aesthetic accelerationism. And they’re all bubbling up and swirling round, and they’re not AI aesthetics. They’re new, interesting, human aesthetics that I think will spill out into the world.

Does this openness help keep things safe as well? Because there’s a lot of discussion about AI image generators being used to generate potentially harmful stuff, whether that’s straightforwardly nasty imagery — gore and violence — or misinformation. How do you stop that from happening?

Yeah, so, it’s amazing. When you put someone’s name on all the pictures they make, they’re much more regimented in how they use it. That helps a lot.

That said, we’ve still had some issues at times where, unfortunately, like, the way that social media works everywhere else, you can make a living by causing outrage, and there’s a motivation for some people to come into the community, pay for privacy, then spend a month trying to create the most outrageous and horrifying shock imagery possible, and then try to publish it on Twitter. Then we have to put our foot on that and say, “That’s not what we’re about; that’s not the type of community we want.”

Whenever we see that, we stomp it out. We ban words if we have to. We’ve collected words for things like photorealistic ultragore, and we’ve banned every word within a mile of that.

What about realistic faces — because that’s another vector for creating misinformation. Does the model generate realistic faces?

It will generate celebrity faces and stuff like that. But we don’t generally — we have a default style and look, and it’s artistic and beautiful, and it’s hard to push [the model] away from that, meaning you can’t really force it to make a deepfake right now. Maybe if you spend 100 hours trying, you can find some right combination of words that makes it look really realistic, but you have to really work hard to make it look like a photo. And personally, I don’t think the world needs more deepfakes, but it does need more beautiful things, so we’re focused toward making everything beautiful and artistic looking.

Where did you get the training data from the model from?

Our training data is pretty much from the same place as everybody else’s — which is pretty much the internet. Pretty much every big AI model just pulls off all the data it can, all the text it can, all the images it can. Scientifically speaking, we’re at an early point in the space, where everyone grabs everything they can, they dump it in a huge file, and they kind of set it on fire to train some huge thing, and no one really knows yet what data in the pile actually matters.

“The entire space has maybe only trained two dozen models like this. So it’s experimental science.”

So, for example, our most recent update made everything look much, much better, and you might think we did that by throwing in a lot of paintings [into the training data]. But we didn’t; we just used the user data based off what people liked making [with the model]. There was no human art put into it. But scientifically speaking, we’re very, very early. The entire space has maybe only trained two dozen models like this. So it’s experimental science.

How much did it cost to train yours?

I would say, training models in this space, I can’t speak about our specific costs, but I can say general things. Training image models is probably around $50,000 every time you do it right now. And you never get it right in one try, so you have to use three tries or 10 tries or 20 tries — and you do need a lot — so it adds up. It is expensive. It’s more than what most universities could spend, but it’s not so expensive that you need a billion dollars or a supercomputer.

The costs will, I’m sure, come down for both training and running. But the cost to run it is actually quite high. Every image costs money. Every image is generated on a $20,000 server, and we have to rent those servers by the minute. I think there’s never been a service for consumers where they’re using thousands of trillions of operations in the course of 15 minutes without thinking about it. Probably by a factor of 10, I’d say it’s more compute than anything your average consumer has touched. It’s actually kind of crazy.

Speaking of training data, one contentious aspect here is the issue of ownership. Current US law says you can’t copyright AI-generated art, but we don’t quite know whether people can assert copyright over images used in training data. Artists and designers work hard to develop a particular style, but what happens if their work can now be copied by AI bots? Have you had many discussions about this?

We do have a lot of artists in the community, and I’d say they’re universally positive about the tool, and they think it’s gonna make them much more productive and improve their lives a lot. And we are constantly talking to them and asking, “Are you okay? Do you feel good about this?” We also do these office hours where I’ll sit on voice for four hours with like 1,000 people and just answer questions.

A lot of the famous artists who use the platform, they’re all saying the same thing, and it’s really interesting. They say, “I feel like Midjourney is an art student, and it has its own style, and when you invoke my name to create an image, it’s like asking an art student to make something inspired by my art. And generally, as an artist, I want people to be inspired by the things that I make.”

But there’s surely a huge self-selection bias at work there because the artists who are active in the Midjourney Discord are bound to be the ones who will be excited by it. What about the people who say, “It’s bullshit; I don’t want my art to be eaten up by these huge machines.” Would you allow these people to remove themselves from your system?

We don’t have a process for that yet, but we’re open to it. So far, I would say it doesn’t have that many artists in it. It’s not that deep of a dataset. And the ones who have made it in have been giving us like “we don’t really feel intimidated by this” answers. Right now, it’s so new; I think it makes sense to play it by ear and be dynamic. So we’re constantly talking to people. And actually, the number one request we get right now from artists is they want it to be better at stealing their styles, so they can use it as part of their art flow even better. And that’s been surprising to me.

It might be different for other [AI image] generators because they try to make something look like the exact thing. But we have more of a default style, so it really does look like an art student being inspired by something else. And the reason we do that is because you always have defaults, so if you say “dog,” we could give you a photo of a dog, but that’s boring. From a human standpoint, why would you want that? Just go to Google image search. So we try to make things look artistic.

That’s something you’ve mentioned a few times in our conversation — the default art style of Midjourney — and I’m really fascinated by this idea that each AI image generator is its own microcosm of culture, with its own preferences and expressions. How would you describe Midjourney’s particular style, and how have you consciously developed it?

[Laughing] It’s a little ad hoc! We try lots of things, and every time we try a new thing, we render out a thousand images. And there’s not really an intention to it. It should look generally beautiful. It should respond to specific things and vague things. We definitely want it to not look like photos. We might make a realistic version at one point, but we wouldn’t want it to be the default. Perfect photos make me a little uncomfortable right now, though I could see legitimate reasons why you might want something more realistic.

I think the style would be a bit whimsical and abstract and weird, and it tends to blend things in ways you might not ask, in ways that are surprising and beautiful. It tends to use a lot of blues and oranges. It has some favorite colors and some favorite faces. If you give it a really vague instruction, it has to go to its favorites. So, we don’t know why it happens, but there’s a particular woman’s face it likes to draw — we don’t know where it comes from, from one of our 12 training datasets — but people just call it “Miss Journey.” And there’s one dude’s face, which is kind of square and imposing, and he also shows up some time, but he doesn’t have a name yet. But it’s like an artist who has their own faces and colors.

Speaking of these sorts of defaults, one big challenge within the image-generation space is dealing with bias. There’s research that shows that if you ask an AI image model to draw a CEO, the CEO is always a white man, and when you ask it to output a nurse, the nurse is always a woman and often a person of color. How have you dealt with that challenge? Is it a big problem for Midjourney or of more concern for corporate companies who want to monetize these systems?

Well, Miss Journey is definitely more of a problem than a feature, and we’re working on something now that will try to break up the faces and give you more variety. But there are downsides of that, too. Like, we had a version where it just completely destroyed Miss Journey, but if you really wanted, say, Arnold Schwarzenegger as Danny DeVito, then it would completely destroy that request [too]. And the tricky thing is getting that to work without wiping out whole genres of expression. Because it’s really easy to have a switch that bumps up diversity, but it’s difficult to have it only turn on when it should.

What I can say is that it’s never been easier to make an image with whatever diversity you want — you just use the word. You’re always one word away from creating, you know — like, I was playing around with “African cyberpunk wizards,” and it looks beautiful, and it’s fucking cool, and all I needed was like one word to tell the model what you want.

So, just to pull back a bit, you’ve talked a lot about how you don’t see the work you’re doing in Midjourney as, shall we say, practical. I mean, it’s obviously very hands-on, but your motivation is more abstract — about the relationship between humans and AI; about how we can use AI in this humanistic way, as you put it. Some people in the AI space tend to think about this technology in the grandest possible terms; they compare it to gods, to sentient life. How do you feel about this?

For a while, I’ve been trying to figure out “what is [Midjourney’s AI image generator]?” Because you can say it’s like an engine for imagination, but there’s something else, too. The first temptation is to look at it through an art lens. To ask: is this like the invention of photography? Because when photograph was invented, paintings got weirder because anybody could take a photo of a face, so why would I paint that picture now?

And is it like that? No, it’s not like that. It’s definitely weirder. Right now, it feels like the invention of an engine: like, you’re making like a bunch of images every minute, and you’re churning along a road of imagination, and it feels good. But if you take one more step into the future, where instead of making four images at a time, you’re making 1,000 or 10,000, it’s different. And one day, I did that: I made 40,000 pictures in a few minutes, and all of a sudden, I had this huge breadth of nature in front of me — all these different creatures and environments — and it took me four hours just to get through it all, and in that process, I felt like I was drowning. I felt like I was a tiny child, looking into the deep end of a pool, like, knowing I couldn’t swim and having this sense of the depth of the water. And all of sudden, [Midjourney] didn’t feel like an engine but like a torrent of water. And it took me a few weeks to process, and I thought about it and thought about it, and I realized that — you know what? — this is actually water.

Right now, people totally misunderstand what AI is. They see it as a tiger. A tiger is dangerous. It might eat me. It’s an adversary. And there’s danger in water, too — you can drown in it — but the danger of a flowing river of water is very different to the danger of a tiger. Water is dangerous, yes, but you can also swim in it, you can make boats, you can dam it and make electricity. Water is dangerous, but it’s also a driver of civilization, and we are better off as humans who know how to live with and work with water. It’s an opportunity. It has no will, it has no spite, and yes, you can drown in it, but that doesn’t mean we should ban water. And when you discover a new source of water, it’s a really good thing.

And Midjourney is a new source of water?

[Laughing] Yeah, that’s a little scary when you say it that way.

I think we, collectively as a species, have discovered a new source of water, and what Midjourney is trying to figure out is, okay, how do we use this for people? How do we teach people to swim? How do we make boats? How do we dam it up? How do we go from people who are scared of drowning to kids in the future who are surfing the wave? We’re making surfboards rather than making water. And I think there’s something profound about that.

Most Popular