OpenAI Dall-E 3 Review: Generative AI for Fanciful, Fun Illustrations

8.0/ 10
SCORE

OpenAI Dall-E 3

Buy at OpenAI

Pros

• Understands long, complex queries
• Engaging, dynamic images
• Conversational style allows easy modifications

Cons

• Photorealistic results can look fake
• Slow to generate images
• Requires $20 a month ChatGPT Plus subscription

With Dall-E, OpenAI helped blaze the trail for generative AI that turns a text prompt into an image. Now there's plenty more competition, but version 3 of the service still holds up.

In my tests comparing it to Adobe Firefly and Google ImageFX, I found Dall-E 3 often did the best job with realistic and engaging images and just about always did the best with surreal fantasies. It's pokey, but it's most likely to give you good, usable results on your first try, especially if you're looking for AI hallucinations that are fun instead of failures.

Dall-E was also the best at encouraging you to get zany and explore what's possible. I'm sure there are designers, artists, programmers and others who are capable of bringing their visions to fruition, but I'm not that skilled. So for me, Dall-E is a marvel.

Dall-E encourages a kind of over-the-top prompt engineering, where people submit paragraphs of text, something between a vignette and a short story -- the kind of prompts that are rejected as too long by some rivals. Check this collective vision of Kansas settlers dreaming of an era of abundance after conquering nature and Native Americans. That's an image generated from a 186-word prompt. It's a form of computer-amplified creativity that's fascinating, and Dall-E is the best tool for the job that I've tried.

Dall-E 3 is available only through the premium ChatGPT Plus service at $20 per month, which also gets you access to a more responsive version of the ChatGPT chatbot and OpenAI's useful GPT Store with custom versions of its AI tools. You can try the earlier Dall-E 2 for free if you want a taste of what's possible, but its results aren't as good.

OpenAI says it may use content submitted to Dall-E 3 to improve the model's performance, that it shares content with a select group of "trusted service providers," and that it doesn't sell data or share content with third parties for marketing. You can also submit a privacy request to have OpenAI stop training on your data or delete your account. See OpenAI's general privacy FAQ and main privacy policy for details.

Here's a closer look at what I found with Dall-E 3.

How CNET tests AI image generators

CNET takes a practical approach to reviewing AI image generators. Our goal is to determine how good it is relative to the competition and which purposes it serves best. To do that, we give the AI prompts based on real-world use cases, such as rendering in a particular style, combining elements into a single image and handling lengthier descriptions. We score the image generators on a 10-point scale that considers factors such as how well images match prompts, creativity of results and response speed. See how we test AI for more.

How good are the images, and how well do they match prompts?

ChatGPT does the best of the text-to-image AI tools I tried when it comes to producing useful, entertaining, believable results. It still makes plenty of errors, like a pickleball player whose paddle sprouts out of his head instead of the paddle grip, but the results made me want to explore more, not close the browser tab. It did a better job with dynamic scenes, with contact and interactions between different subjects, and with moods.

ChatGPT is an instrumental part of Dall-E. It magnifies your prompts, adding florid prose to inject drama into the results. It also makes for a conversational usage style: you can ask for an image, then ask for a tweak without having to resubmit the entire query.

ChatGPT's language technology is what enables it to process long, elaborate prompts, too. It turns out that advanced word-handling abilities are useful for advanced image-handling abilities.

That helps Dall-E 3 outperform rivals including Adobe's Firefly and Google's ImageFX when it comes to turning your prompt into what you want, assembling multiple elements correctly. For example, Dall-E 3 was the only AI image generator I tested that successfully created a dragon flying over a castle, breathing fire and clutching a white fluffy sheep in its talons. Granted, it was cradling the sheep gently, perhaps in response to OpenAI's strictures against violence, but it was close.

Don't expect perfection. Dall-E bungled plenty of details. In an image of a dog walker overwhelmed by too many dogs, the human is hilariously overmatched by his dozens of canines. But if you look closely, common AI problems are visible: one of the dogs had two heads, another was a cat, and others had disturbing problems with legs, ears and tongues. But the image is still compelling.

Another problem I had was its inability to create lifelike images. Dall-E 3 is inclined to create dreamy scenes, often with a style that's more an illustration than a photorealistic scene. Personally, I'm fine with that, since photorealism often reveals the most cringeworthy shortcomings of AI imagery. But photorealism is an important use case, and I had a hard time trying to guide Dall-E toward it.

I never could get a satisfactory image of a British Navy sea captain on a sailing ship. Dall-E couldn't figure out how people hold telescopes, didn't understand what a sextant is at all, and in one case populated a ship with 12 captains and no crew. The gist of what I wanted was often there, but it wasn't put together convincingly.

Dall-E 3, like its rivals, can't count. It did correctly show a single pool ball when I asked for that, but it was an 8 ball that was half green and half white. That's not something you'll see at the local billiards hall.

ChatGPT Plus subscribers also get access to at least 10 custom logo generators GPTs based on Dall-E that are tuned for that task and available on the GPT Store.

Again and again, images were marred with detailed problems like this. But also again and again, I still liked the results despite that. Exploding balloons, biblically accurate angels, red-tailed hawks, coffee shop logos -- nothing was ever perfect, but often it was good enough. 

Perhaps we'll all collectively evolve a new set of aesthetic rules to accommodate AI images, accepting some wrongness in exchange for their value. If so, Dall-E 3 is going to do well.

How engaging are the images?

Very engaging. Over and over, Dall-E 3 produced vibrant, attention-getting images. Even when there were problems, I often enjoyed them. They made me chuckle at times and look at details.

Dall-E 3's maximalist language approach can be unwanted at times, though. When prompting for an image of a doctor and patient surrounded by medical equipment, there were a dozen monitors tracing out heartbeat and respiration data. One of the computers had about 100 keys on the keyboard.

People can look somewhat crazed with emotion too. My request to render a frustrated person standing behind a box of cleaning supplies produced a couple people who looked more enraged than frustrated and one who was downright demonic.

You can ask Dall-E 3 to tone things down, and sometimes it will.

Can you fine-tune results?

The text-based interface of Dall-E 3 is conversational. Unlike Adobe's Firefly, there are no buttons for image styles or parameters. You can get used to its conversational style, but this longtime image editing software user likes buttons and sliders.

You can ask for images to be widescreen, portrait or landscape, and the AI will oblige. But when you start with a fresh image prompt, it'll sometimes revert to its square default. More than once, I got a square image I liked, but you can't simply ask to expand that exact image. (You can with Photoshop's generative expand feature, though, if you want to go that route.)

How fast do images arrive?

Good things come to those who wait, I guess. Dall-E 3 often took 20 or 30 seconds to produce just one image. That often exceeded my patience, so I usually headed over to my email inbox for a couple minutes then came back to check the results.

That lag can put a drag on the back-and-forth interactivity of the ChatGPT style of operation. But I'll take slow speed and good results over fast response and inferior images.

Generative AI pushes computing technology to its limits. OpenAI has learned to squeeze better results out of ChatGPT, so I hope it'll bring similar efficiencies to Dall-E.

Conclusion

Dall-E 3 is an impressive tool that can inject some creative fun into your life and do useful image creation work. As with all text-to-image generation tools, it's prone to errors, but Dall-E 3 offers the best results among rivals in my testing. You'll have to decide for yourself if the relative quality -- and the best version of the ChatGPT chatbot -- is worth $20 a month of your budget.

Editors' note: CNET is using an AI engine to help create a handful of stories. Reviews of AI products like this, just like CNET's other hands-on reviews, are written by our human team of in-house experts. For more, see CNET's AI policy and how we test AI.