How to Use the Dall-E 3 AI Art Generator to Create Stunning Images From Text (2024)

When we first wrote this story, OpenAI’s Dall-E 2 was one of the major names in generative AI art, along with a few others like Midjourney and Stable Diffusion. Since that long-ago era (called 2022), Dall-E has come a long way. We're now on the third iteration (Dall-E 3), which boasts significant improvements, particularly in understanding text prompts and more precision (in theory) thanks to on-the-fly prompt customization.

The image creator is also now fully integrated with ChatGPT Plus, the $20 per month paid version of OpenAI's text AI generator. Dall-E 3 is likewise part of ChatGPT Team and the corporate-focused ChatGPT Enterprise.

The use of Dall-E 2 is no longer an option. If you had art stored with it, it remains accessible for now. Grab the good stuff while you can.

Dall-E 3 (along with all generative AI) use comes with several ethical concerns since the capabilities of most are built on the backs of hardworking artists who never gave permission for their work to be used in training. One "fix" for that now built into Dall-E is that it won't always mimic specific requested styles, especially if there's a copyright question. Also, artists can ask to opt out their images from training "future image generation models," according to the website.

Dall-E 3 is built into other tools. For example, users of the Shutterstock stock art service can use it to generate new images and know they're being a little more ethically prudent. (Shutterstock art was, after all, used in part to train the Dall-E AI.) Microsoft's big investment in OpenAI put ChatGPT 4/Dall-E 3 in place as the generators of choice for its various Copilot options built into Windows and Office. (Copilot implementation also has ethics issues.)

How, exactly, do you work with Dall-E 3? Is it truly 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 a perfect result.

Sign Up and Pay Up

Start with an account for the standard ChatGPT at chat.openai.com. If you don't have one, use an Apple, Google, or Microsoft account, or create a login with an email address and astrong password. There’s no multifactor authentication option at this stage.

Once you're in, you'll see Upgrade Plan at the lower left, or the menu at the top that says ChatGPT 3.5, which offers a choice to switch to GPT-4/Upgrade to Plus. Either will show this screen, with the Upgrade to Plus option.

(Credit: OpenAI)

The upgrade actually costs $20 plus tax; at 8% that is $21.60 per month. It clearly says, "You'll be charged the amount and at the frequency listed above until you cancel." There's no trial option. After you pay and sign in, the desktop interface looks the same, but be sure to select ChatGPT 4 from the menu at the top. If you click your name at the bottom and check settings, you can turn on things like multifactor authentication now.

The menu at the top will also inform you that under GPT-4 and with Dall-E 3, you'll get 40 messages within three hours. That's down from the 50 messages within three hours it offered at launch. The reduction was likely made to keep traffic in check as the use of the service grows.

Prep for the Chat

Unlike the now-defunct Dall-E 2, which had its own interface, Dall-E 3 is just another aspect of ChatGPT 4, where you send a "message" and the chatbot responds. Thankfully, with GPT-4 you can have an ongoing conversation to change/adapt/improve the things it generates for you. That includes images. On the premium ChatGPT Plus, prompts are limited to 4,000 characters—approximately 500 words.

Note that Dall-E 3 also works in the ChatGPT app for mobile (available for iOS and Android). Simply log in with the same ChatGPT Plus account. With the mobile interface, you'll get options for using a picture taken with your device's camera, uploading a pre-existing image, or uploading a file. There's also a headphones icon that activates the option to use your voice for input.

ChatGPT interface on iOS (Credit: OpenAI)

Since we're talking about creating images, you should start your prompt with "Create an image."

Dall-E 3 can use an image of your own as the basis for new AI-generated content. To upload a picture, use the paper clip attachment icon.

To make art with generative AI, the prompt is the thing. It's where most people stumble. An initial prompt just tossed out to a generative art AI tends to make something...interesting. It's seldom perfect. Things are always a little mushy. A little odd. A little off. The image turns out a lot better once you perfect the prompt.

The Perfect Prompt

Design promptshave bothcontentsubject terms (what you want to see) andstyle 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. You can also use several other modifiers to boost quality, focus Dall-E on particular things, or just insert some extra creativity. Use emotional words to get a more positive or negative image and set the mood or aesthetic. While you can ask for things in certain styles, some requests will trip up ChatGPT as it attempts to follow the rules of copyright (see below).

The more specific your details and the more you use proper art or photography terminology, the better. Denote camera shot info like the camera model, focal length, etc. to get better faux photos, for example.

You'll notice when an image is churned out that ChatGPT Plus has made some iterative changes to your prompt to improve output. Mention in the prompt "do not add details" or "do not modify/alter the prompt" if you feel your wording is enough to get what you want. If you find yourself doing it all the time with every new prompt, click your name at the lower right and select Customize GPT to set up a modifier that will force the chatbot to use your prompts without changes.

Not feeling very creative? Hire people to create the right prompt to get the right art. But if you’re going to do this a lot, getting the prompt right upfront all by yourself is a boon.

The biggest helper of all might be ChatGPT itself. You can ask it to write you a prompt for Dall-E with a few basic words (like "amazing giant robot"). It'll spit out a blurb full of details you may not have considered. Some call this "promptception."

Using ChatGPT Plus to make prompt to feed into Dall-E 3. (Credit: OpenAI/Eric Griffith)

It's up to you to then refine the prompt more with further messages. Do things like ask for a change in the following:

• Point of view

• Number of subjects

• Aspect ratio

• Add or remove colors

A new prompt for Dall-E 3 makes a new image; it won't make simple adjustments to the existing image. Sometimes it tosses out your original entirely. Make sure to ask for the Generation ID for any image you like (see below).

Click the Regenerate icon (it looks like an "undo" arrow circling left) below an image to make a new version. It'll keep the old versions. Keep regenerating as much as you want, but each click probably eats into your message per hour allotment with ChatGPT and the button seems to disappear after a certain amount of time.

If you really don't like an image, click the thumbs down to indicate the result as a "bad response." This elicits some questions from ChatGPT.

Bad Response options on a Dall-E 3 image (Credit: OpenAI/Eric Griffith)

Dall-E 3 has two style options. The default is vivid for that hyperreal feeling it is particularly good at. Try typing "Style: natural" in the prompt to get a more realistic-looking image. It's worth experimenting with, though vivid vs. natural may be in the eye of the beholder. Likewise, there are two quality options: standard (the default) or HD for high-definition to get more details. Specify which you want in a prompt.

Changing Sizes and Retaining Styles

With Dall-E 3, you only get one image per prompt by default. However, you can specify that it generates as many as 10 from one prompt. (With Dall-E 2 you got four variations from which to pick. Same with Microsoft's Copilot; again, Copilot is powered by Dall-E 3.) Or you can enter multiple unique prompts in the chat to receive multiple images, each one exclusive to its prompt, not all variations on the same wording.

Images from Dall-E arrive by default in a square measuring 1024x1024 pixels by default. However, you can specify it to be taller (1792x1024) or wider (1024x1792) by using those numbers in the prompt or say "portrait orientation" or "landscape orientation."

What if you want a larger-scale version of the same image? There are third-party tools to do it (see below), but Dall-E 3 can enlarge pics, too. Simply say something like "Upscale this image 2x using code interpreter." Those last three words are key—they force Dall-E 3 to simply enhance the existing image rather than regenerate it. You won't see the upscaled image again in the chat; you'll receive a download link instead.

A download link for an upscaled image in Dall-E 3. (Credit: OpenAI)

If you want to duplicate an image's style later with different content, ask ChatGPT for the "gen_ids" (short for Generation ID) for the most recent image. Every pic made by Dall-E gets one. It'll give you a long string of numbers to copy such as "yRmG5bW4bmcfWbVP." You can also use the Generation ID to go back and have Dall-E 3 regenerate a variation on an image even after the regenerate icon disappears.

Later, enter a prompt and append the words "in the style of image ______." Or say "Give me image _______ but change the clothes to something formal." Whatever you want. Enter the string of numbers for the blank.

Saving the Gen ID for any image you like is important for future iterations of the image. Here's a great guide to using Gen ID.

Can you generate a consistent character/persona to use repeatedly in different Dall-E 3 images, like you can with Midjourney? Almost. If you ask for the "seed" for the image, you'll get a small string of numbers, which you can then use to try and recreate the image with some consistency from the previous image. (You can also assign a seed number to the image in your initial prompt, just pick a few digits.) Next, try a prompt such as "modify the image with seed ### but change the background."

A seed number is also a way to recreate an image if you have the exact prompt used previously. You could also create a GPT for yourself that helps maintain some regularity image-to-image. But the consistency is unlikely to ever be 100%. It's just now how generative AI works for now.

If you access Customize ChatGPT under "How would you Like ChatGPT to respond," enter "when outputting Dall-E 3 images, always display the Gen_ID and Seed." You'll see this metadata every time in the future.

Customization to ChatGPT Plus can output metadata with every Dall-E 3 image. (Credit: Dall-E 3/Eric Griffith)

Don't forget that ChatGPT Plus also supports GPTs, custom chatbots that run with specific parameters. You'll find them in the left-hand navigation under Explore GPTs. There are several of them featured that are specific to Dall-E 3 image generation but many more in the search. You can also make your own.

What Dall-E 3 Won't Do

OpenAI has a policy about not depicting things like hate symbols, violence, deceptive topics, or directly showing anything that's obviously under copyright. Nor will it generate images of real people. Well, it's not supposed to, anyway.

In the past, you could reference films and shows to approximate their looks and style, but when I asked Dall-E 3 to make a Star Wars-esque robot, ChatGPT chastised me, saying "Creating an image that is 'Star Wars'-esque falls under the restrictions of our content policy, which prevents me from generating images directly inspired by copyrighted material." I took that part out and got a clunky, artistic droid anyway. You can see the difference below from 2022 Dall-E 2 vs 2024 Dall-E 3.

The prompt 'Star wars droid making a painting of digital flowers on to a stretched canvas, over the shoulder, photo-realistic, 1970 film style, mid-day lighting through a window' generated both image in different years. (Credit: Dall-E 2/Dall-E3/Eric Griffith)

Later, I asked it to render an image I uploaded in the style of The Simpsons. It had no problem complying.

With Dall-E 2, users had the option to do outpainting—a feature that would take the initial square image and allow the user to extend the borders, which would fill in with even more AI-generated art. It's similar to the Generative Fill function in Photoshop.

Unfortunately, you can't currently outpaint natively with Dall-E 3 in ChatGPT. At best, you can say, "Revamp this image into portrait mode" or something similar, but the pic will still max out at 1792x1024. (The Dall-E 2 robot above at top was partially generated via outpainting, as seen here.)

Outpainting in Dall-2: Generating a frame to extend a Dall-E 2 image. (Credit: Dall-E 2/Dall-E 3/PCMag)

When creating images with AI, avoid asking it to reproduce text in the image when possible. At least keep your expectations low. In the ultimate irony, despite being part of a text-generating chatbot, Dall-E and other AI art generators like Midjourney don't actually know anything about text. Be incredibly specific. If there's text you want in the image, put it in quotation marks in the prompt. Still, don't be surprised if any text rendered in your AI image is utter gibberish. For now.

An example of Dall-E 3's inability to handle text. (Credit: Dall-E 3/Eric Griffith)

Inpainting is Back

The old Dall-E 2 offeredinpainting, letting you delete part of an image with the eraser tool and replace it with something AI-generated based on a prompt. You could even do it with an uploaded image.

While there is currently still no native outpainting in Dall-E 3, limited inpainting is back.

On the desktop click any generated image to go full screen with the editor interface—apparently, OpenAI isn't going to call it 'inpainting' anymore. The chat will move to the right side.

At the top, you'll see an icon of a swirling paintbrush. It gives you access to a resizable selection tool. Select the areas you'd like to change. Enter a prompt indicating what you want changed, and it'll happen without mucking up the rest of the generated image. As you can see below I went from a cat in the snow to a pitbull pup in the same snow, on the run from flying saucers—but kept the other details like trees and lighting.

(Credit: ChatGPT/Eric Griffith)

The selection tool is optional; you can simply use a prompt to change a cat to a dog. But to keep it limited to an area of the image you must specify that in words, or you risk having Dall-E 3 generate an entirely new image.

You can also give orders like "make this image monochrome,"—but even that will often create a new image. To keep the old, the Selection tool is a must. For example, with that exact prompt, I got this instead, with what is clearly a copyright violation on Ridley Scott's Alien xenomorph.

Unintended results abound in Dall-E 3. (Credit: Dall-E 3/Eric Griffith)

The interface for the inpainting function is similar on the mobile version of ChatGPT. Click an image generated by Dall-E 3 to go into the editor. From there, you can select Edit to enter a prompt or click the Select paintbrush to highlight just part of an image you want to change. You can undo or redo changes before it finalizes the update.

Importantly, inpainting/editing is not available on images that you upload.

Downloading and Editing

When you're ready to take ownership of an image you made with Dall-E 3, click the down arrow icon in the upper right corner of the image and you'll get the option to download a WebP version. Or right-click to select Save Image As in your browser.

WebP is the new default image type, but it's a type many users don't want to work with. However, you can always right-click a Dall-E 3 image and copy it to paste into an image editor to save it in another format. We also recommend the Save Image as Type browser extension for Chrome or Edge, which lets you pick the image format (JPG, PNG, or WebP) to download with a right-click.

If you want to upscale your image resolution after the fact, you can also try a website likebigjpg.com, ARC Face Restorer(which also fixes faces), or commercial software like the $99 Gigapixel AI for macOS or Windows. Of course, these tools also work on images you create with Dall-E 3 elsewhere, like with Copilot, or any other generative AI, like Midjourney or Google's Gemini.

You can also upscale images in the Microsoft Designer—part of the Copilot generative AI powered by Dall-E 3. Using that is easiest if you make the images in Copilot.

Done with Dall-E?

If you've tried out the paid version of ChatGPT Plus to get to Dall-E 3 and found it all wanting—or you'd prefer to hire real people (applause!)—you can always cancel your $20 subscription. Click your name at the lower left corner and go to My Plan > Manage My Subscription > Cancel Plan.

Prompt: 'Show an over-the-shoulder image of a person at a computer canceling their subscription to ChatGPT Plus.' (Credit: OpenAI's Dall-E 3/Eric Griffith)