Select text to annotate, Click play in YouTube to begin
00:00:00
Seven years ago, back in 2015, 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?
00:00:26
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.
00:00:54
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.
00:01:24
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.
00:01:56
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.
00:02:27
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.
00:02:52
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.
00:03:18
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.
00:03:45
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.
00:04:12
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.”
00:04:44
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...
00:05:10
...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
00:05:35
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.
00:06:08
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.
00:06:35
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.
00:07:00
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.
00:07:28
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.
00:07:52
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.
00:08:22
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
00:08:48
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.
00:09:16
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.
00:09:44
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.”
00:10:21
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
00:10:45
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.
00:11:13
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.
00:11:40
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
00:12:08
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,
00:12:35
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.
00:13:05
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. And they had a lot of interesting things to say. So I've compiled them into a bonus video. Please check it out and add your own thoughts in the comments. Thank you for watching.
End of transcript