one major development in AI research  was automated image captioning.

Machine learning algorithms could  already label objects in images,  

and now they learned to put those labels  into natural language descriptions.

And it made one group of researchers curious.

What if you flipped that process around?

We could do image to text.

Why not try doing text to  images and see how it works?

It was a more difficult task.They didn't want  

to retrieve existing images  the way google search does.

They wanted to generate entirely novel scenes that didn't happen in the real world.

So they asked their computer model for something it would have never seen before.

Like all the school buses you've seen are yellow.

But if you write “the red or green school bus”  would it actually try to generate something green?

And it did that.

It was a 32 by 32 tiny image.

And then all you could see is like a  blob of something on top of something.

They tried some other prompts like “A herd  of elephants flying in the blue skies”.

“A vintage photo of a cat.”

“A toilet seat sits open in the grass field.”

And “a bowl of bananas is on the table.”

Maybe not something to hang on your wall  but the 2016 paper from those researchers  

showed the potential for what might  become possible in the future.

And uh... the future has arrived.

It is almost impossible to overstate how far  the technology has come in just one year.

By leaps and bounds. Leaps and bounds.

Yeah, it's been quite dramatic.

I don't know anyone who  hasn't immediately been like

“What is this? What is happening here?”

Could I say like watching waves crashing?

Party hat guy.

Seafoam dreams.

A coral reef. Cubism.

Caterpillar.

A dancing taco.

My prompt is Salvador Dali painting  the skyline of New York City.

You may be thinking, wait  AI-generated images aren't new.

You probably heard about this generated portrait  going for over $400,000 at auction back in 2018.

Or this installation of morphing portraits,  which Sotheby's sold the following year.

It was created by Mario Klingemann, who  explained to me that that type of AI  

art required him to collect a specific dataset of  images and train his own model to mimic that data.

Let's say, Oh, I want to create landscapes,  so I collect a lot of landscape images.

I want to create portraits,  I trained on portraits.

But then the portrait model would not  really be able to create landscapes.

Same with those hyper realistic  fake faces that have been plaguing  

linkedin and facebook – those come from a  model that only knows how to make faces.

Generating a scene from any combination of words  requires a different, newer, bigger approach.

Now we kind of have these huge  models, which are so huge that  

somebody like me actually cannot train  them anymore on their own computer.

But once they are there, they are  really kind of— they contain everything.

I mean, to a certain extent.

What this means is that we can now  create images without having to actually  

execute them with paint or  cameras or pen tools or code.

The input is just a simple line of text.

I'll get to how this tech works later in the video  

but to understand how we got here,  we have to rewind to January 2021

When a major AI company called Open AI announced  DALL-E – which they named after these guys.

They said it could create images from text  captions for a wide range of concepts.

They recently announced DALLE-2, which promises  more realistic results and seamless editing.

But they haven't released  either version to the public.

So over the past year, a community of  independent, open-source developers  

built text-to-image generators out of other  pre-trained models that they did have access to.

And you can play with those online for free.

Some of those developers are now working  for a company called Midjourney, 

which created a Discord community with bots that  turn your text into images in less than a minute.

Having basically no barrier to entry to  this has made it like a whole new ballgame.

I've been up until like two  or three in the morning.

Just really trying to change things, piece things together.

I've done about 7,000 images. It's ridiculous.

MidJourney currently has a wait-list for  subscriptions, but we got a chance to try it out.

"Go ahead and take a look."

“Oh wow. That is so cool”

“It has some work to do. I feel like it can  be — it's not dancing and it could be better.”

The craft of communicating  with these deep learning  

models has been dubbed “prompt engineering”.

What I love about prompting  for me, it's kind of really  

that has something like magic where you have to  know the right words for that, for the spell.

You realize that you can refine  the way you talk to the machine.

It becomes a kind of a dialog.

You can say like “octane render blender 3D”.

Made with Unreal Engine...

...certain types of film lenses and cameras...

...1950s, 1960s...

...dates are really good.

...lino cut or wood cut...

Coming up with funny pairings, like a Faberge Egg McMuffin.

A monochromatic infographic poster about  typography depicting Chinese characters.

Some of the most striking images  can come from prompting the model  

to synthesize a long list of concepts.

It's kind of like it's having a very strange  collaborator to bounce ideas off of and get  

unpredictable ideas back.

I love that!

My prompt was "chasing seafoam dreams,"

which is a lyric from the Ted Leo and the Pharmacists' song "Biomusicology."

Can I use this as the album cover for my first album? "Absolutely."

Alright.

For an image generator to be able to  respond to so many different prompts, 

it needs a massive, diverse training dataset.

Like hundreds of millions of images scraped from  the internet, along with their text descriptions.

Those captions come from things like the alt text  that website owners upload with their images,  

for accessibility and for search engines.

So that's how the engineers  get these giant datasets.

But then what do the models actually do with them?

We might assume that when  we give them a text prompt,  

like “a banana inside a snow globe from 1960."

They search through the training data  to find related images and then copy  

over some of those pixels. But  that's not what's happening.

The new generated image doesn't  come from the training data,  

it comes from the “latent space”  of the deep learning model.

That'll make sense in a minute, first  let's look at how the model learns.

If I gave you these images and told you to match  them to these captions, you'd have no problem.

But what about now, this is  what images look like to a  

machine just pixel values for red green and blue.

You'd just have to make a guess, and  that's what the computer does too at first.

But then you could go through  thousands of rounds of this  

and never figure out how to get better at it.

Whereas a computer can eventually figure out a  method that works- that's what deep learning does.

In order to understand that this arrangement  of pixels is a banana, and this arrangement  

of pixels is a balloon, it looks for metrics that  help separate these images in mathematical space.

So how about color? If we measure  the amount of yellow in the image,  

that would put the banana over here and the  balloon over here in this one-dimensional space.

But then what if we run into this:

Now our yellowness metric isn't very  good at separating bananas from balloons.

We need a different variable.

Let's add an axis for roundness.

Now we've got a two dimensional space with the  round balloons up here and the banana down here.

But if we look at more data we may come  across a banana that's pretty round,  

and a balloon that isn't.

So maybe there's some way to measure shininess.

Balloons usually have a shiny spot.

Now we have a three dimensional space.

And ideally, when we get a new image we  can measure those 3 variables and see  

whether it falls in the banana region  or the balloon region of the space.

But what if we want our model to recognize,  

not just bananas and balloons,  but…all these other things.

Yellowness, roundness, and shininess don't  capture what's distinct about these objects.

That's what deep learning algorithms do  as they go through all the training data.

They find variables that help improve their  performance on the task and in the process,  

they build out a mathematical space  with way more than 3 dimensions.

We are incapable of picturing multidimensional  space, but midjourney's model offered this and I like it.

So we'll say this represents the latent space of  the model. And It has more than 500 dimensions.

Those 500 axes represent variables that  humans wouldn't even recognize or have  

names for but the result is that  the space has meaningful clusters:

A region that captures the essence of banana-ness.

A region that represents the textures  and colors of photos from the 1960s.

An area for snow and an area for globes  and snowglobes somewhere in between.

Any point in this space can be thought  of as the recipe for a possible image.

The text prompt is what navigates us to that  location. But then there's one more step.

Translating a point in that mathematical  space into an actual image involves a  

generative process called diffusion.  It starts with just noise and then,  

over a series of iterations, arranges pixels  into a composition that makes sense to humans.

Because of some randomness in the process,  

it will never return exactly the  same image for the same prompt.

And if you enter the prompt into a  different model designed by different  

people and trained on different  data, you'll get a different result.

Because you're in a different latent space.

No way. That is so cool. What the heck? The brush  strokes, the color palette. That's fascinating.

I wish I could like — I mean he's dead,  but go up to him and be like, "Look what I have!"

Oh that's pretty cool. Probably the  only Dali that I could afford anyways.”

The ability of deep learning to extract  patterns from data means that you can copy an  

artist's style without copying their images,  just by putting their name in the prompt.

James Gurney is an American illustrator who  

became a popular reference for  users of text to image models.

I asked him what kind of norms he would like  to see as prompting becomes widespread.

I think it's only fair to  people looking at this work  

that they should know what the prompt  was and also what software was used.

Also I think the artists should be allowed  to opt in or opt out of having their work  

that they worked so hard on by hand be used  as a dataset for creating this other artwork.

James Gurney, I think he was a  great example of being someone  

who was open to it, started  talking with the artists.

But I also heard of other artists  who got actually extremely upset.

The copyright questions regarding  the images that go into training the  

models and the images that come out  of them…are completely unresolved.

And those aren't the only questions  that this technology will provoke.

The latent space of these models contains some  

dark corners that get scarier as  outputs become photorealistic.

It also holds an untold number  of associations that we wouldn't  

teach our children but that  it learned from the internet.

If you ask an image of the CEO,  it's like an old white guy.

If you ask for images of  nurses, they're all like women.

We don't know exactly what's in the  datasets used by OpenAI or Midjourney.

But we know the internet is biased toward  the English language and western concepts,  

with whole cultures not represented at all.

In one open-sourced dataset,  

the word “asian” is represented first  and foremost by an avalanche of porn.

It really is just sort of an infinitely complex  mirror held up to our society and what we  

deemed worthy enough to, you know, put  on the internet in the first place and  

how we think about what we do put up.

But what makes this technology so  unique is that it enables any of  

us to direct the machine to  imagine what we want it to see.

Party hat guy, space invader, caterpillar, and a ramen bowl.

Prompting removes the obstacles between ideas  and images, and eventually videos, animations,  

and whole virtual worlds.

We are on a voyage here, that  is it's a bigger deal than  

than just like one decade or the  immediate technical consequences.

It's a change in the way humans imagine,  communicate, work with their own culture  

And that will have long range,  good and bad consequences that we  

we are just by definition, not going to  be capable of completely anticipating.

Over the course of researching this video I spoke to a bunch of creative people

who have played with these tools.

And I asked them what they think this all means for people who make a living making images.

The human artists and illustrators and designers and stock photographers out there.