How to Use the Dall-E AI Art Generator to Create Stunning Images From Text

AI art generators have been in the news a lot this year, be it for their amazing advances or questionable uses. OpenAI’s Dall-E 2 is one of the major names in this space. It’s now open to the public and developers, and soon it’ll be built into Microsoft software and the Bing search engine.

Stock art service Shutterstock is also going to integrate the tool, and pay artists it copies to give back and (maybe) avoid some of the ethical concerns. Shutterstock art was, after all, used in part to train the Dall-E AI.

But how, exactly, do you work with Dall-E? Is it really as simple as typing in a description—called a prompt—and getting back a picture? To be honest, yes. But there’s a lot more to be aware of if you want to get anywhere close to the perfect result. Let’s talk about it.

Sign Up and Pay Up

The first thing to do is create an account at labs.openai.com(Opens in a new window). Use a Google or Microsoft account, or create a login with an email address and a strong password. There’s no multifactor authentication option.

Dall-E is not entirely free. The service runs on “credits(Opens in a new window).” You get 50 free credits at signup, and then 15 credits free per month after that, but they don’t roll over. Paid credits do roll over month to month, for up to 12 months; get 115 credits for $15.

One credit lets you do one AI art generation (you get four new images per typical generation). That could start from a prompt, but also a credit is used for creating a variation of already-generated art. You can use a lot of credits trying to drill down on the right AI-generated picture.

The Perfect Prompt

Once signed up, you’ll be confronted with a form for your prompt. If you click the Surprise Me button, you’ll see other random prompts dropped into the text box—they don’t count against your credits until you click Generate. You can also upload an image of your own and use Dall-E to edit it to add new AI-generated content, or to make completely new variations of the original.

(Credit: PCMag)

But the prompt is the thing. It’s likely where most people will stumble. We’ve done our own experiments to see how random prompts can do with Dall-E and its competition like Midjourney(Opens in a new window) to create art. The initial prompt always makes something…interesting. But it’s seldom perfect. Things are always a little mushy. A little odd. A little off. The art turns out a lot better if you perfect the prompt, which is limited to 400 characters (emojis(Opens in a new window) work(Opens in a new window)).

Design prompts(Opens in a new window) have both content (what you want to see) and modifiers (how it should look). For example: “A robot drawing a painting at an easel” is content, but “over-the-shoulder view, colorful, oil paint, in the style of Van Gogh” are all modifiers.

Not feeling very creative? There are people who will help you create the right prompt to get the right art. They go by job titles like Prompt Designers or Prompt Engineers, and their expertise doesn’t come for free. But if you’re going to do this a lot, getting the prompt right up front can save you money, because you won’t use as many credits.

In addition to videos like the one above, there are sources like The Dall-E 2 Prompt Book(Opens in a new window) from dall-ery gall-ery. It’s worth a read. It spells out many “hacks” you should take into consideration in your prompt, such as using terms like “close-up” and possible camera angles, types of lighting, listing eras to mimic, such as “1920s,” or even mentioning a specific camera lens or smartphone type that “shoot” the AI image.

Using emotional words to get a more positive or negative image is big, as are using words to set the mood/aesthetics(Opens in a new window). Prompts to get the most out of AI-generated “photos” and “illustrations” is a whole world in and of itself. The possibilities are endless. Your only limit is 400 characters.

The one thing you can’t do with Dall-E is generate images of real people. Well, you’re not supposed to, anyway. It has filters and limitations in place to make sure, but with billions of images studied, who knows when Dall-E might issue a picture that looks like a celeb. That said, you can reference films and shows to approximate their looks and style.

Can you prompt for a style that mimics a living artist? Of course, and if Dall-E has seen that artist’s work it can probably deliver an imitation. A style can’t be copyrighted, but of course, it’s a crappy thing to do if you’re going to use it commercially when you could have worked with an artist.

Image Variations

Any image you generate in Dall-E—or any image you upload to Dall-E (make sure you’re the copyright holder)—can get an instant variation. Uploaded images have to be cropped down to a square 1:1 ratio image.

(Credit: PCMag)

You can also create variations of a variation. Just keep clicking on the three-dot menu on images and select Variation to make more.

Variations can’t really be controlled. It’s simply Dall-E looking at the content of an image and creating something approximately the same, with the same styles, similar layouts, etc.

Let’s say you’ve created an image with Dall-E and you like it. Mostly. But there’s something not quite right. Select Edit, and delete the part you don’t like with the eraser tool, re-write part of the prompt to address that section (say a face that needs more specific description, or a whole new background to replace the original). Dall-E creates a variation based on the new specs in the prompt. Your prompt still has to describe the entire desired image, not only the erased area.

This eraser option also works with images you can upload, so you could delete a head or a whole background. Write a prompt for what you want to appear in the erased section. For example, erase some sky and use the prompt “an upside-down crop-duster airplane” to see it appear overhead. Working within the confines of a generated image in Dall-E is called “inpainting.”

Erase part of your upload, create a prompt, and generate new AI art in that spot.
(Credit: PCMag)

Make an Image Larger

So what is “outpainting” then? Well, generally the output from Dall-E is limited to you downloading an image measuring 1,024-by-1,024 pixels. Another thing you can do under Edit is to create Generation Frames. Click the Add Generation Frame icon that looks like a box with a plus sign on the upper left (or press the letter F), and you’ll get a floating box you can place  anywhere outside the image perimeter.

Generating a frame to extend a Dall-E image.
(Credit: PCMag)

Click Generate again and the image will extend into that frame, as if the AI had just continued drawing/shooting into the extended space with the same prompt. Keep doing it until the image is as big as you want. Each new frame generated uses up a credit, of course.

When you’re ready to own it, click the download button (the down arrow) and you’ll get a .PNG file of your new masterpiece. If you really want to upscale your image resolution after that, you can try a website like bigjpg.com(Opens in a new window) or ARC Face Restorer(Opens in a new window) (which also fixes faces) or a commercial piece of software like the $99.99 Gigapixel AI(Opens in a new window).

Remember, you can always go back to your My Collection page on the Dall-E site and see all of the images you’ve ever created, including variations. So if you had an idea to improve an older image you never downloaded, you can get back to it.

If you created something really amazing, you can click Share > Publish. This puts the image on a public page at openai.com, for which you get a special link(Opens in a new window) to share with others. It will show off the image plus the prompt used to create it. If you don’t want to go that public, try Save > Favorites to put it in a collection that you can view later.

Become an Artist

Ready to try your hand at AI artistry? All you need to do is sign up and type. Give it a try at Dall-E with the free credits and you’ll have a taste for all the AI-generated art that is likely to come our way in the next several decades.

A variation on an upload, extended with new frames.
(Generated on Dall-E)