March 2023 - Michael's Programming Bytes

An intro to DALLE 2

We’ll start our journey down the AI art rabbit hole by first discussing the basics of DALLE 2.

First of all, what is DALLE 2? Well, DALLE 2 is the second version of the DALLE AI art algorithm-both DALLE and DALLE 2 were created by OpenAI-the same lab that created the ChatGPT chatbot. In fact, ChatGPT and both iterations of the DALLE art algorithm utilize the GPT (or Generative Pre-Transformer) NLP neural network. The original iteration of DALLE was released in January 2021 and DALLE 2 was released as a beta test in July 2022.

Setting up DALLE 2

Now, how would you start using DALLE 2? First of all, click on this link to be navigated to DALLE’s homepage-https://openai.com/product/dall-e-2:

Once you get to the DALLE 2 homepage, click on the Try DALLE link to start working with DALLE 2. Once you click on this link, you’ll need to sign up for a free DALLE 2 account (if you have a Gmail account, you can simply use these credentials for signing up).

After signing up for a DALLE 2 account, you’ll see a screen that looks like this:

Once you see this screen, you can type a prompt into the text box and have fun creating AI art!

If you haven’t figured out where the name DALLE 2 comes from, it’s simply a portmanteau of the artist Salvador Dali and the PIXAR robot WALL-E (which was a great movie by the way).
You only get 15 free DALLE 2 prompts a month, so use them wisely. Of course, you can always pay for more prompts if you feel inclined to do so-the cheapest deal is 115 prompts for $15 (not bad).

And now let’s create some AI art!

Let’s start with a simple DALLE 2 prompt-perhaps A coloring book page featuring two tabby cats and a ball of yarn. Here’s the output we get:

As you can see, DALLE 2 does quite a great job of creating a coloring book page featuring two tabby cats and a ball of yarn-it even returns one partially colored page.

This prompt, along with any other prompt you type into the input box, will generate four AI art images based off of what you typed into the input box.

Now, let’s try this prompt-A painting of a cow jumping over the moon in the style of Andy Warhol-and see what kind of output we get:

As you can see, DALLE 2 did quite a good job of creating an Andy Warhol-style painting of a cow jumping over the moon. If you’re familiar with Warhol’s work, you’ll be amazed at how well the DALLE 2 algorithm replicates his style (though I would be remiss not to note that DALLE 2’s amazing ability to recreate any art style isn’t without controversy, as this algorithm can easily mimic many art styles without the consent of the artists).

AI art still has a long ways to go

Look, DALLE 2 is smart enough (or rather, built well enough) to generate thousands and thousands of images by working its deep learning magic to mimic thousands of different art styles. But-at least as of March 2023-AI art is still far from perfect.

Let’s say we wanted to generate an AI image of an ice cream shop with the sign Mike's Ice Cream Shop with DALLE 2. Here’s what happens:

In this example, I used the prompt A colored pencil sketch of an ice cream store with a sign that says "Mike's Ice Cream Shop". The four art pieces that were generated create great colored-pencil sketches of an ice cream shops, but none of the storefront signs say “Mike’s Ice Cream Shop”, which is part of the request I sent to DALLE 2. Rather, all the signs generated contain gibberish text (my favorite one is the second picture which has a sign reading “Mik Mic Shke”).

OK, so DALLE 2 can’t really generate a good sign for my made-up ice cream store, but can it generate a good logo for this blog? Let’s find out:

OK, so I asked DALLE 2 to generate a logo for this blog and include the blog’s name-Michael’s Programming Bytes-and its slogan-“Byte sized programming classes for all coding learners”-on the logo. Much like the “Mike’s Ice Cream Shop” example above, the four logos generated don’t contain either the blog’s name or slogan. What the AI-genreated logos contain, however, is a lot of gibberish (though if I ever created a blog called “Mtheglyles” with the slogan “Byilyse”, I’d certainly use the first AI-generated logo).

So, we can see that DALLE 2 isn’t so good with inserting string of text into its AI-generated art. However, can DALLE 2 generate images of people? Let’s find out:

In this example, I typed in the prompt A watercolor painting of President Joe Biden and as you can see from the lack of output above, my request was denied by DALLE 2.

If you click the content policy hyperlink, you’ll be redirected to DALLE 2’s content creation policy, which would give you a better idea as to why this request was denied:

As you can see from the content policy screenshot above, this request was denied because DALLE 2 generally doesn’t accept politically-themed prompts and my Joe Biden prompt fell into that category.

Now, let’s try another prompt that contains a public figure, but this time, let’s make it a non-political public figure. Take a look at the prompt below:

In this example, I used the prompt A photo of a movie poster with Ryan Reynolds' face on it and DALLE 2 generated four images of movie posters with what it thinks is Ryan Reynolds’ face on it. Granted, just like with the ice cream store example, the text on these AI-generated movie posters is pure gibberish, but the face on the first AI-generated poster does resemble Ryan Reynolds pretty closely. The face on the second poster bears somewhat of a resemblance to Reynolds while the third face (and especially the fourth face) looks almost nothing like him.

Interestingly enough, when I swapped Ryan Reynolds’ name (but left the rest of the prompt unchanged) for a female actress-Gal Gadot-this is what I got:

My best guess as to why DALLE 2 will generate images of some public figures (without 100% accuracy) and not others is, aside from their content policy, that OpenAI (the lab that makes DALLE 2) doesn’t want it to be too easy for people to make deepfakes-which in this day and age, would be a fair reason to make it hard to generate fully-accurate images of public figures.

Now that we’ve discovered the limits of DALLE 2 when it comes to generating images of public figures, let’s see how this algorithm does when it comes to generating images of general people:

In this example, I used the prompt A photo of friends on a college campus and the AI-generated results have been quite hit-or-miss. Granted, DALLE 2 did a good job of generating the background-a generic college campus in this case-but DALLE 2 didn’t quite have the same magic when it came to generating images of the people. Let’s take a closer look at one of those images (to take a closer look at an image, simply click on it):

As you can see in this image above, DALLE 2 did a great job of generating the background-a generic college campus-but didn’t do such a good job of generating the generic college students (especially the students’ faces). Also, if you zoom into this picture really closely, you’ll see that the young lady in the orange tank-top has six fingers on one hand.

Yes, even AI has its biases

Aside from DALLE 2’s sometimes imperfect art generations, another thing to note about this algorithm (and AI at large) is that, just like humans, AI has its biases too.

Are you familiar with unconscious bias? If not, it’s a phenomenon that affects how you behave around other people based off of assumptions and/or beliefs you may have about other people just based off of appearances (e.g. baggy clothes, skin color, etc.) rather than their character.

Well, AI does have its unconscious biases too. Think about it-who do you need to create and maintain AI infrastructure? Humans! Since humans have their unconscious biases, they can often incorporate their biases into the programs that they create (though I’m sure this isn’t true for every programmer/developer).

Let’s observe AI bias in action through this DALLE 2 prompt:

So the prompt I used-An oil painting of an American rapper-seems quirky enough, right? Well, take a look at the four AI-generated images and tell me what they have in common. Since all the AI-generated paintings are of black men, this does look like a clear example of unconscious bias in AI.

Let’s try one more example:

OK, so I used the prompt A photo of a kindergarten teacher and surprisingly, got less biased photos than I did in the previous example (but still, 3/4 of the AI-generated photos are of women, and not a ton of diversity in the AI-generated photos).

Interestingly enough, AI seems to fine generating images of people when it’s a single person. When there are multiple people in the picture (as you saw with our college friends example), things fall apart.

Before I go

Before we go, I just want to leave you with some final things I wanted to mention about DALLE 2.

To download any image generated, click on the image itself and once the down arrow icon appears, click on it to download the image:

Also, as per a 2022 ruling by the U.S. Copyright Office, AI-generated images don’t have any copyright protections on them (yet), which means you can use the images as freely as you’d like, but you also can’t claim a copyright on any images you generate through DALLE 2, since after all, the art technically wasn’t created by you but rather by a bunch of 1s and 0s.

Thanks for reading,

Michael

Month: March 2023

And Now Let’s Create Some AI Art! (AI pt.6)