Skip to main content

Everything You Wanted to Know About Dall-E 2, the Amazing AI Artist

Tim Arends has a Bachelor of Arts degree in Fine Art from Berea College, KY. He has written on a variety of technological subjects.

The Hanging Gardens of Babylon in the middle of a city, in the style of Dalí (Actual text prompt used)

The Hanging Gardens of Babylon in the middle of a city, in the style of Dalí (Actual text prompt used)

The Artists Formerly Known as Human

There’s a brand new artist in the world, and not only is he going to revolutionize the entire field of art, he’s going to radically alter the fields of publishing, Internet creation and even creativity itself! He’s going to change the way we think about art, intelligence, imagination and even what it means to be human!

The artist I’m talking about is not a person, it’s a machine. A machine that will create paintings for you!

Now wait a minute, before you scoff and say, “That’s nothing new; Photoshop has had filters that make any photograph look like a painting for years now,” that’s not what I’m talking about. I’m talking about a new artificial intelligence that will create a painting for you from scratch!

Again, you scoff.

But wait. Hear me out. This will create a painting for you from scratch in any style imaginable, from comic books to Renaissance painting.

You scoff for a third time. No, you do not have to provide it with a source image. Like I said, it will create it from scratch!

But there's more. To use this AI, you don’t have to be a programming expert; you don’t have to tweak a bunch of knobs and dials and settings and parameters to get just what you want. You simply tell it what you want it to draw or paint. And it will do it for you.

Simply enter a plaintext phrase describing what you want: a cottage, a rock festival, a robot insurrection, or pigs flying through the sky. And it will give you exactly what you want, it will do it in any style you desire, and it will look indistinguishable from a drawing or painting created by a talented professional with years of experience and training.

A song of neon light, digital art (actual text prompt used)

A song of neon light, digital art (actual text prompt used)

Dall-E 2 From the OpenAI Project

Dall-E 2 is so new that I hadn’t heard about it until about a week before starting this essay (and I try to keep abreast of issues relating to art and artificial intelligence).

No, Dall-E 2 doesn't work by simply finding images that match part of the text prompt and copying and pasting that part of the image into a new image. I tested it by doing a Google image search on several Dall-E 2 images and I could find nothing that looks even remotely similar to any of the images or parts of the image that Dall-E 2 created. Clearly, Dall-E 2 creates its artwork by understanding the concept or concepts involved in whatever image it is creating well enough to paint it in exquisite detail, from any angle and in virtually any artistic style.

And incredible as it is, as of this writing, most people still haven’t heard of this technological innovation, and even if they have, most haven’t begun to grasp the tremendous changes it’s soon going to make on all of society. For example:

  • What will this technology do to the art world? It’s already difficult to make a living as an artist. Will this technology make it impossible?
  • If this technology can create photorealistic images of any fantastic, fictional concept we can imagine, will we be able to trust photography or photojournalism anymore?
  • If this technology can understand exactly what we ask of it and can give it to us in a way that’s even better than we imagined, can it be said that this technology has achieved human level creativity or greater? Has it achieved human level intelligence? If it can understand the emotion we are attempting to infuse into an image and can give it to us, can it be said to understand human emotions, or even to share our emotions?

These are just a few of the mind boggling questions this new technology will raise. So in this essay, let’s attempt to answer a few of them.

Portrait of a shiba inu astronaut, oil painting, 16th century (actual text prompt used)

Portrait of a shiba inu astronaut, oil painting, 16th century (actual text prompt used)

What Is Dall-E 2 Exactly?

Dall-E 2 is an artificial intelligence system developed by OpenAI, a research and development company. Dall-E 2 understands objects and it understands the relationship between objects. Therefore, it can generate an image of a koala bear, as well as a koala bear riding a motorcycle and wearing a funny hat.

It is a vast improvement on OpenAI’s previous project, Dall-E 1. The name is simply a play on the name of the artist Salvador Dali and the robot Wall-E from the Pixar/Disney movie of the same name.

The astonishing thing is how quickly Dall-E has been improving, with the vastly superior Dall-E 2 coming out just one year after Dall-E 1.

I tested this text-to-image engine back in October of 2020. Back then text-to-image was little more than a novelty (the above image is supposed to show "a man"). Look how far it has come in just over one year!

I tested this text-to-image engine back in October of 2020. Back then text-to-image was little more than a novelty (the above image is supposed to show "a man"). Look how far it has come in just over one year!

How Does Dall-E 2 Work?

You simply tell it what you want and in what style by typing a short phrase, and it paints it for you.

On a deeper level, Dall-E and other image generation AIs do not create artwork the same way a person does. When you create a drawing, you probably sketch it out first in a rough form. Then you may concentrate on the face of the person you’re drawing, filling out the details of the eyes, nose and mouth, and work from there.

Artificial intelligence works in a way that is alien to you and me. Rather than focusing on one part of the image and working out from there, it concentrates on drawing the whole image all at once! Through the process of diffusion, it starts with a pattern of noise and then over a series of iterations transforms those pixels into a recognizable object.

It starts by filling the whole image area with a shade of gray, which develops into blurry shapes that gradually become sharper as it goes along. (In reality, it lays down a pattern of pixels which are so fine they look to the human eye like gray. Then it removes whatever pixels do not look like the image it has decided to create!)

Another thing that may or may not be different between the way humans and AI works: when you start a drawing, you probably have a particular idea in your mind of how it should look, but as you work, you come up with little inspirations and your idea takes various twists and turns as you go along so that the finished product may be much different from what you originally envisioned.

In contrast, I used to think the AI knows exactly what it wants to do the moment it starts and never changes its mind once during the process, but then I watched the online art engine Synth.Run work, which gives you a preview of what it’s doing as it goes along, and it seems to change its mind several times as it works, flashing a number of very different images until it finally settles on one.

Amazingly, the AI is said not to look at the images the way we do, but merely as pixel values, represented by binary ones and zeros spread across the entire image. In training, it randomly attempted to match these sets of data to the appropriate caption. As it got more and more matches correctly, it learned to identify the pixel values with the image description.

An astronaut sits within a futurist cockpit overlooking Jupiter (actual text prompt used)

An astronaut sits within a futurist cockpit overlooking Jupiter (actual text prompt used)

How Was Dall-E 2 Developed?

According to a video by Vox, it all started in 2015 with automated image captioning. You know how you can do a Google image search and it will find images matching your description, even if those images were not labeled by a human?

Pinterest has a similar feature, in which you can circle a single object in an existing image and it will attempt to identify the object and find other pictures that contain that object.

The way the developers trained the AI was with a massive data set of captioned images—hundreds of millions of images scraped from the Internet along with their text descriptions. Conveniently, many or most images on the Internet are already provided with alt descriptions for the visually impaired and for search engine optimization.

The next challenge researchers tackled was getting the AI to name all the objects in a picture, and then to write a simple sentence or phrase describing that image.

The next logical step after this? Flipping the process, or getting the AI to draw an image based on a phrase. Obviously, this is a much more difficult task, but the researchers eventually succeeded in generating very small, lo-res and blurry pictures.

The amazing thing is how quickly the AI improved, going from small, blurry thumbnail images to the incredibly detailed, artistic and creative images generated by Dall-E 2. OpenAI announced Dall-E in January 2021 and the new and improved version, Dall-E 2, only about a year later.

Another ingenious way in which the AI was trained was through the use of GANs, or generative adversarial networks. Essentially, an AI image creation engine was pitted against a recognition engine. The creation engine tried to create images so realistic that they fooled the recognition engine, which was programmed to try to spot the fakes. At first, the creation engine had a very low rate of success, but over time its success rate got better and better to the point where it fooled the recognition engine most of the time!

Creating systems in which the AI trains itself was a major key to its fantastic rate of advance because computers work at a super-fast speed, they never tire and they never need to rest.

A monkey head that is only made out of fruit, 3D (Actual text prompt used)

A monkey head that is only made out of fruit, 3D (Actual text prompt used)

Why Is Dall-E 2 So Revolutionary?

Dall-E 2 (and similar technologies) are revolutionary in more ways than can be counted. It understands concepts and relationships, such as animals, sitting, riding, man-made objects, vehicles, and so on, and you can combine them, so It can draw a koala bear wearing a hat riding a skateboard. It’s revolutionary because of the number of styles it can produce, from cartoony to children’s storybook to photorealism. It’s revolutionary in its ability to understand natural language requests. It’s revolutionary in the sheer quality of its artwork, often indistinguishable from that created by a professional artist.

And it’s revolutionary in how quickly it has been improving.

Avocados dancing, drinking, singing and partying at a Hawaiian luau (Actual text prompt used)

Avocados dancing, drinking, singing and partying at a Hawaiian luau (Actual text prompt used)

Can I Use Dall-E 2 Right Now?

No, Dall-E 2 has not yet been released to the general public. You can sign up for a waiting list on the website, however. You can also admire the Dall-E 2 art posted to the Dall-E 2 Instagram account and that of the select few people who currently have access to Dall-E 2.

A highly detailed, vintage, grainy photo of flying cars in New York City (1936) - Actual text prompt used

A highly detailed, vintage, grainy photo of flying cars in New York City (1936) - Actual text prompt used

Are There Other Competing Technologies Out There?

Yes, the field of natural language art generation has exploded, with a number of competing technologies arriving seemingly overnight. Independent open source developers built a bevy of text to image generators out of existing pre-trained models (although, in terms of sophistication, Dall-E 2 is still at the head of the pack, as of this writing). In the chart below is a very incomplete list composed by myself with the help of Google.

Just a few months ago, an AI that could create elaborate works of art from a simple text prompt was virtually unheard of; now it’s practically everywhere. Developers have even created a Discord community populated with text-to-image bots!

Keep in mind that generating artwork from a text prompt takes a lot of computing power, so these services run on their own online servers, not on your computer or mobile device. Therefore, you need an Internet connection to use them.

Neon psychedelic blacklight sunset (Actual text prompt used)

Neon psychedelic blacklight sunset (Actual text prompt used)

Competing AI Projects Similar to Dall-E

Service/AppMonetizationOwnership

StarryAI

Freemium

Proprietary

Dream by WOMBO

Free

Proprietary

Synth.run

Free

Proprietary

Craiyon

Free

Open Source

DALL-E FLOW in Google Colab

Free

Open Source

MindsEye Beta

Free

Open Source

NeuroGen

Freemium

Proprietary

ruDALL-E

Free

Open Source

VQGAN+CLIP

Free

Open Source

NeuralBlender

Freemium

Proprietary

StarryAI: Winter Chilled Spring Bud Heaven (Actual text prompt used)

StarryAI: Winter Chilled Spring Bud Heaven (Actual text prompt used)

What Can I Use Image Generators Like Dall-E 2 for?

It depends on which image generator you’re talking about. If it’s Dall-E 2, you can presumably use it to create any image you would otherwise use clip art, stock images or even a professional artist for, but Dall-E 2 has not yet been released to the general public.

However, there are other image generators that are not quite up to the level of Dall-E 2 (yet) that are nevertheless good enough to use for a wide variety of applications (including illustrating essays on this and the rest of the Hubpages family of websites). Best of all, you can use them right now!

Following is just a short list of applications that many of the current crop of image generators are good enough to create images for:

  • Calendars
  • Websites
  • Informational articles
  • Essays about science and technology
  • Illustrating your own short stories
  • Fantasy-style images (think wizards and warriors)
  • Greeting cards
  • Posters
  • Music albums
  • Cups, mugs, t-shirts, etc.
  • Illustrating Bible passages
  • Church and school presentations
  • Science fiction-style images
  • Political images and memes
  • Children’s stories
  • Art for posting on Instagram and other image sharing websites
  • Spice up your social media presence
  • Inspiration or starting point for your own artwork
  • Imagining how cities will look in the future (as demonstrated by SFGate)
StarryAI: Elizabethan Conjurer (Actual text prompt used) (Rococo style)

StarryAI: Elizabethan Conjurer (Actual text prompt used) (Rococo style)

More Ways to Use AI Generated Art

  • Take your own favorite novel (especially in the fantasy or science fiction genre) and plug some of the scenes from it into the image generator. Voila! You’ve just created an illustrated version of your favorite novel!
  • Use image generators to illustrate some of the classics of literature. Stories like Grimm's Fairytales would lend themselves well to illustration by image generators.
  • Write your own fan fiction about one of your favorite movies, games or TV shows and illustrate it with AI-generated artwork.
  • An image generator can help you visualize passages from the Bible. Many Bible scenes come to life when illustrated (possibly for the first time) in vivid detail and in full color.
  • Many people, I believe, will use image generators to expand their own imaginations. Already, some use them first thing in the morning for inspiration with which to start the day.
  • I find image generators a fascinating exploration into the mind of artificial intelligence, how exactly it works and what it is capable of. It’s almost like having a visual conversation with a friendly alien life form!
StarryAI: Noah building the ark (Actual text prompt used)

StarryAI: Noah building the ark (Actual text prompt used)

What About NFT Artwork?

If you are a traditional artist, you may never have heard of the term NFT artwork, and the reason is that it is not a traditional art term. It is a term relating to cryptocurrency.

Some marketers of art generation services make a big deal out of using their products for generating NFT artwork. For example, such a use is listed at the very top of the webpage and Apple App Store listing for StarryAI.

NFT stands for non-fungible tokens, and the explanation of exactly how they work can be as complicated as that of cryptocurrency itself. At its most basic level, however, an NFT is simply unique digital content that can be bought and sold on a blockchain, such as on the Ethereum platform. The virtual content can take many forms, such as music, art or virtual trading cards.

NFT marketplace platforms such as OpenSea, Nifty, Gateway and Minitable have sprung up. The NFT process ensures that the buyer is purchasing full ownership of a digital asset, such as a piece of artwork.

However, those who already have a significant online following are the most likely to succeed in selling NFT artwork. Just because you generate a piece of artwork and put it up for sale does not mean anyone will want to buy it! Some people argue that NFTs are a fad whose heyday has already passed.

Even if you never get involved in selling NFTs, you can still use AI-generated art for any of the purposes listed above.

StarryAI: The art deco emerald city of Oz fame (Actual text prompt used)

StarryAI: The art deco emerald city of Oz fame (Actual text prompt used)

Who Owns the Rights to AI-Created Images?

The exact rights conferred depend on the service that is used, but according to the Starryai FAQ, “You are the copyright owner of your creations as long as you have the right to use all assets used in generating the creation.”

StarryAI gives the option to upload an image to use as a starting point (but it is not necessary), and as long as you own the rights to the starting image, if any, you own the rights to the resulting artwork.

The US Copyright Office says an AI can’t copyright its own art, but that only applies when the AI claims to be the sole creator, with no human intervention whatsoever. If a human has a hand in creating the artwork, even if it is simply providing a text prompt or pressing a button, then the art is considered copyrighted by the human who initiated the process.

StarryAI: Artificial intelligence will take over the world! (Actual text prompt used)

StarryAI: Artificial intelligence will take over the world! (Actual text prompt used)

How Can a Human Claim to Have Created It?

This is a difficult—and philosophical—question to answer. Some people believe that if an artificial intelligence creates an image, it removes a human from the process entirely, so he or she can in no way claim to have created the image. Likewise, many people believe that if you use a filter on a photograph which makes the photograph look like a painting, that can also not be considered a form of artwork.

My own mother disliked computer-generated artwork, such as the use of filters on photographs. She considered it a form of “cheating,” giving the person who created it an unfair advantage over other artists who did it “the hard way”—by hand with a brush and paint.

This question is as old as art itself. Who can say what is art and what is not? As noted above, legally, the person who initiates the artwork creation process with an AI is considered the “artist” and the owner of the work, but that is different from whether he or she is considered the artist in the creative, non-legal sense.

Even artists themselves can’t agree with each other on this issue. There is a difference between the art of Michelangelo and that of Picasso, yet both are considered by the art world today to be equally gifted and respected artists.

When the camera was invented, it created a question as to whether artists would be needed any longer. Indeed, it probably caused the loss of many jobs that would otherwise have gone to artists. Every print advertisement, for example, that uses a photograph rather than a drawing or painting is a job that went to a photographer rather than a traditional artist.

The invention of the camera helped to spur a revolution in the field of art and the development of whole new artistic schools. Many artists went to a more abstract form of artwork with the rationale that it was no longer needed to depict something literally or exactly, since that is what the camera was for.

Even some photography, such as the work of Ansel Adams, has gained grudging acceptance as a form of art. Yet there has always been a battle with the general public over what is and is not considered “art.” Is splashing paint on a piece of canvas art? How about turning a crucifix upside down and suspending it in a vial of urine? These forms of “art” have all spurred controversy.

There is also a difference between fine art and commercial art. Fine art, generally speaking, is art that hangs on people’s walls and in art galleries. Commercial art is usually used to sell a product, whether it be in a print advertisement or whether the artwork itself is part of the product, as in an animated film or comic book. Fine artists often look down their noses at commercial art, just as the general public often looks down its noses at the more abstract examples of fine art.

Ultimately, I think, it is the buyer who determines whether something is art or not, whether that be an art gallery or someone who wants to use artwork to sell a product. The real question is whether there are going to be that many buyers of art when an AI makes the art creation process so easy that virtually anyone can become an “artist.”

The crazy, far out world of artificial intelligence (Actual text prompt used)

The crazy, far out world of artificial intelligence (Actual text prompt used)

If I Use DALL-E 2 Art, Should I Give It Full Credit?

Should you admit upfront that you use an image generator to create the artwork that you post online? Or should you allow or encourage the assumption that you did it yourself, by hand, using traditional art techniques or media? After all, you’re legally considered the artist, so why should you have to give the credit to an image generator?

This is a prickly question. Before the general public becomes aware of artificial intelligence image creation, they will likely credit you with great artistic skill when you announce that you are the creator of a particular AI-generated image.

However, when they become aware of AI image generators, there will probably be a big letdown and, human nature being what it is, they will consider you to have "cheated" and not to be the "true" creator of the art at all.

As time goes on, and image generators become ubiquitous, probably most people will use them to generate images for their Facebook accounts, their Instagram pages and so on.

I can even see the day when most people automatically assume that any new piece of art they see online was created at least partially by an artificial intelligence. Or, maybe they will assume the more mediocre pieces of art to have been done by hand and the more sophisticated ones to have been done by an AI!

In the meantime, it's not a bad idea, I think, to mention the name of the service that you use to generate the artwork. In other words, name yourself as the artist, but name DALL-E or StarryAI or whatever as the "tool" that you used. This way, you are engaging in full disclosure and cannot be accused of trying to "put something over" on your viewers.

Will AI Art Destroy the Jobs of Human Artists?

Ah, this question gets into the deep philosophical and practical issues raised by AI-generated artwork and the incredibly profound implications such artwork has for the entire field of art…and even for humanity in general!

For a deep dive into all of these issues, be sure to see my companion essay in which I explore the monumental effects I think this technology will soon have on all of society!

This article is accurate and true to the best of the author’s knowledge. Content is for informational or entertainment purposes only and does not substitute for personal counsel or professional advice in business, financial, legal, or technical matters.

© 2022 Timothy Arends