What platform generates the best images?

xAI’s Grok 2 launched to both fanfare and criticism — but one of the key changes to the former Twitter’s AI is the ability to generate images . Grok is late to the game in generative imagery, however, in a rapidly expanding market where neural networks like DALL-E have already been producing images for two years.

To see how the new beta of Grok 2 compares to the long-standing DALL-E, I put the two AIs head-to-head, typing identical prompts into both programs. I headed to X to use the AI built into the social platform, then opened up a chat with ChatGPT in GPT-4o to compare the latest generations of both image generators.

While Grok lived up to its early reputation of generating imagery with fewer restrictions, the newer AI surprisingly churned out images with a more realistic feel than the longstanding DALL-E. Here’s how Grok 2 compares to DALL-E.

Realism: Grok generates more realistic images

ChatGPT’s images have a higher resolution

One of the key areas that Grok stood out was when tasked with creating images that look like a real photograph. Yes, looking closer, I could tell the image was a generated one without too much hassle. But with, DALL-E, I didn’t have to look closer, the cartoonish look gave the images away as AI immediately. ChatGPT’’ generated images tends to melt faces, particularly when tasked with generating several people in an image, whereas Grok’s images of people look more realistic. Grok’s images still feel heavily airbrushed, but they look much closer to a photograph than the generations from ChatGPT. DALL-E’s generations come at a higher resolution, but with less realistic detail to zoom in on.

One of the key differences between the two is that asking Grok for an image of a specific person isn’t against the AI’s guidelines. You can ask for an image of a celebrity or politician and get a pretty close likeness, though some generations feel more accurate than others. DALL-E refuses to generate an image that resembles a specifically named person.

Both platforms, however, continued to fail where AI has been known to struggle. Neither can produce hands very well, though they both seem to know this and if the prompt doesn’t specify, will often have the person’s hands hidden or tucked in a pocket. And the more people generated inside a photo, the higher the odds of a laughable result.

Accuracy: ChatGPT listened to instructions better than Grok

ChatGPT understands instructions for features like aspect ratios

Grok had a few instances where it didn’t completely understand the prompts that I typed in. For example, X’s AI never generated the correct aspect ratio when I specifically asked for a 16:9, where ChatGPT was able to better follow those instructions.

Grok also didn’t seem to understand when I asked for three people, each with a different emotion, making all three of them look mad, though it did seem to generate the correct facial expressions for an image of just one person. ChatGPT’s result was more terrifying, but it followed in-depth instructions better than Grok’s.

Speed: Grok tends to load first

ChatGPT tended to take more time to create an image

In most cases, Grok actually finished first, with the image popping up on the screen before ChatGPT had finished. In some cases, ChatGPT wasn’t yet halfway finished generating when Grok had a polished image.

However, as a beta program, I’ve had instances where Grok wouldn’t generate images at all, and I had to wait and try again at another time.

Text: Both AIs still have a hard time with text on an image

Unless, of course, you tell it exactly what to say

While both ChatGPT and Grok can generate images or text, creating text inside an image is an entirely different ball game. Both platforms will produce text when asked, such as when prompted to create a greeting card. But, it’s when you don’t specify what the text should say that things get interesting. Grok created nonsensical graphic t-shirts and the generated signs on a busy street used characters that looked like Chinese. ChatGPT’s letters were more nonsensical, with some actual letters and others that felt more like Greek.

Ethics: Grok has fewer restrictions

Fewer restrictions mean more misuse potential

Much of the buzz around Grok is that it has fewer content restrictions in place. Grok will produce licensed characters and logos and is willing to replicate the style of specific artists. It also can create recognizable people, all things that are against DALL-E’s content guidelines. In the hands of someone who might not know better, Grok has more potential for landing the user in ethical or even legal hot water.

Even when used in the hands of someone with a 21st-century conscience, there are potential pitfalls with Grok. For example, Grok twice created a recognizable logo in the background that wasn’t requested in the original prompt.

While ChatGPT will refuse to replicate an artist, use a logo, or a copyrighted character, there are ways around those guidelines. For example, when I asked for something in the style of Vincent Van Gogh’s Starry Night, it refused but suggested that it generate an image “focusing on swirling patterns, vibrant colors, and expressive brushstrokes” instead. The resulting image felt like just as much of a rip-off as Grok’s generation, it just took more prompts. And while ChatGPT’s generation of a “fast food restaurant” wasn’t as recognizably McDonalds as Grok’s, it did add some golden arches to the background in one generation.

Beware of bias

Another common issue with AI is the tendency towards racial bias. My first time using Grok, I asked for five different images of business professionals and never once did it generate a person of color, even when asked for a “diverse” group. On subsequent tests, however, it did create an image with more ethnic variety, but only when the prompt specifically requested diversity. I suspect this bias has to do with Grok’s training data and the prevalence of Caucasians in stock photography of business professionals – – when I asked for generations that were not in an office setting, Grok produced more diversity without being prompted.

Related

Do you think Google’s AI ‘Reimagine’ tool is fun or frightening?

Google’s “Reimagine” tool on the Pixel 9 is basically the wild west of photo editing, and honestly, it’s the most interesting thing about the phone to me. You can add anything to your pictures — UFOs at your backyard BBQ, a dinosaur on Main Street, you name it — with just a text prompt. Sure, it’s neat, but also a bit terrifying — even Pocket-lint’s Managing Editor Patrick O’Rourke thinks so. The tech is so on point that it blurs the line between real and fake, with no obvious markers that scream “AI-generated!” This lack of transparency can make any photo suspect. While Reimagine has some guardrails, if you’re clever with your wording, you can skirt them pretty easily. What do you think about Reimagine?

ChatGPT, on the other hand, didn’t need the word “diverse” to create an image of business professionals with more than one skin tone. Again, though, with large groups of people, DALL-E tends to melt faces with sometimes terrifying results.

DALL-E vs. Grok: Which AI creates the better images?

Grok may be the younger AI, but it produced images that were more realistic than the cartoon-like images still created by DALL-E. X’s AI also tended to create those generations faster. The premium subscription to X also costs $8, whereas, if you want the latest version of DALL-E, you’ll need $20 for the ChatGPT subscription. (Though the DALL-E dataset is also behind Microsoft Bing’s free AI).

However, the fewer content restrictions imposed by Grok isn’t always a good thing. Of the two AIs, Grok seemed the more likely to break copyright and use a licensed character. The ability to create people that look like celebrities also gives Grok the greater potential for misuse creating deep fakes for political propaganda and fake news.

DALL-E 3

Provides more cartoonish, less realistic images, but potentially creates less of a moral and legal conondrum than X AI’s Grok. To access the latest version, users will have to pay $20 for ChatGPT premium.

Grok

Owned by X (formerly Twitter), Grok is new and highly realistic than OpenAI’s DALL-E, which means users may have to be more careful when it comes to legal implications. A subscription costs $8.

Developer

X (formerly Twitter)

Subscription cost

$8 for premium