What platform generates the very best photographs?

xAI’s Grok 2 launched to each fanfare and criticism — however one of many key modifications to the former Twitter’s AI is the power to generate images . Grok is late to the sport in generative imagery, nonetheless, in a quickly increasing market the place neural networks like DALL-E have already been producing photographs for 2 years.

To see how the brand new beta of Grok 2 compares to the long-standing DALL-E, I put the 2 AIs head-to-head, typing an identical prompts into each packages. I headed to X to make use of the AI constructed into the social platform, then opened up a chat with ChatGPT in GPT-4o to match the newest generations of each picture mills.

Whereas Grok lived as much as its early status of producing imagery with fewer restrictions, the newer AI surprisingly churned out photographs with a extra lifelike really feel than the longstanding DALL-E. This is how Grok 2 compares to DALL-E.

Realism: Grok generates extra lifelike photographs

ChatGPT’s photographs have the next decision

One of many key areas that Grok stood out was when tasked with creating photographs that appear like an actual {photograph}. Sure, wanting nearer, I might inform the picture was a generated one with out an excessive amount of trouble. However with, DALL-E, I didn’t must look nearer, the cartoonish look gave the photographs away as AI instantly. ChatGPT’’ generated photographs tends to soften faces, significantly when tasked with producing a number of individuals in a picture, whereas Grok’s photographs of individuals look extra lifelike. Grok’s photographs nonetheless really feel closely airbrushed, however they appear a lot nearer to {a photograph} than the generations from ChatGPT. DALL-E’s generations come at the next decision, however with much less lifelike element to zoom in on.

One of many key variations between the 2 is that asking Grok for a picture of a selected individual is not in opposition to the AI’s tips. You possibly can ask for a picture of a celeb or politician and get a reasonably shut likeness, although some generations really feel extra correct than others. DALL-E refuses to generate a picture that resembles a particularly named individual.

Each platforms, nonetheless, continued to fail the place AI has been identified to battle. Neither can produce fingers very effectively, although they each appear to know this and if the immediate does not specify, will usually have the individual’s fingers hidden or tucked in a pocket. And the extra individuals generated inside a photograph, the upper the chances of a laughable consequence.

Accuracy: ChatGPT listened to directions higher than Grok

ChatGPT understands directions for options like side ratios

Grok had just a few situations the place it did not fully perceive the prompts that I typed in. For instance, X’s AI by no means generated the proper side ratio once I particularly requested for a 16:9, the place ChatGPT was in a position to higher comply with these directions.

Grok additionally didn’t appear to know once I requested for 3 individuals, every with a special emotion, making all three of them look mad, although it did appear to generate the proper facial expressions for a picture of only one individual. ChatGPT’s consequence was extra terrifying, but it surely adopted in-depth directions higher than Grok’s.

Velocity: Grok tends to load first

ChatGPT tended to take extra time to create a picture

Generally, Grok truly completed first, with the picture popping up on the display screen earlier than ChatGPT had completed. In some circumstances, ChatGPT wasn’t but midway completed producing when Grok had a elegant picture.

Nonetheless, as a beta program, I’ve had situations the place Grok would not generate photographs in any respect, and I needed to wait and take a look at once more at one other time.

Textual content: Each AIs nonetheless have a tough time with textual content on a picture

Except, after all, you inform it precisely what to say

Whereas each ChatGPT and Grok can generate photographs or textual content, creating textual content inside a picture is a completely completely different ball sport. Each platforms will produce textual content when requested, comparable to when prompted to create a greeting card. However, it’s whenever you don’t specify what the textual content ought to say that issues get fascinating. Grok created nonsensical graphic t-shirts and the generated indicators on a busy avenue used characters that seemed like Chinese language. ChatGPT’s letters had been extra nonsensical, with some precise letters and others that felt extra like Greek.

Ethics: Grok has fewer restrictions

Fewer restrictions imply extra misuse potential

A lot of the excitement round Grok is that it has fewer content material restrictions in place. Grok will produce licensed characters and logos and is prepared to duplicate the type of particular artists. It can also create recognizable individuals, all issues which might be in opposition to DALL-E’s content material tips. Within the fingers of somebody who may not know higher, Grok has extra potential for touchdown the person in moral and even authorized sizzling water.

Even when used within the fingers of somebody with a Twenty first-century conscience, there are potential pitfalls with Grok. For instance, Grok twice created a recognizable brand within the background that wasn’t requested within the unique immediate.

Whereas ChatGPT will refuse to duplicate an artist, use a brand, or a copyrighted character, there are methods round these tips. For instance, once I requested for one thing within the type of Vincent Van Gogh’s Starry Evening, it refused however prompt that it generate a picture “specializing in swirling patterns, vibrant colours, and expressive brushstrokes” as a substitute. The ensuing picture felt like simply as a lot of a rip-off as Grok’s technology, it simply took extra prompts. And whereas ChatGPT’s technology of a “quick meals restaurant” wasn’t as recognizably McDonalds as Grok’s, it did add some golden arches to the background in a single technology.

Watch out for bias

One other widespread problem with AI is the tendency in direction of racial bias. My first time utilizing Grok, I requested for 5 completely different photographs of enterprise professionals and by no means as soon as did it generate an individual of colour, even when requested for a “numerous” group. On subsequent checks, nonetheless, it did create a picture with extra ethnic selection, however solely when the immediate particularly requested range. I think this bias has to do with Grok’s coaching knowledge and the prevalence of Caucasians in inventory images of enterprise professionals – – once I requested for generations that weren’t in an workplace setting, Grok produced extra range with out being prompted.

Associated

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

Google’s “Reimagine” software on the Pixel 9 is principally the wild west of photograph modifying, and truthfully, it’s essentially the most fascinating factor concerning the cellphone to me. You possibly can add something to your footage — UFOs at your yard BBQ, a dinosaur on Principal Road, you identify it — with only a textual content immediate. Certain, it is neat, but in addition a bit terrifying — even Pocket-lint’s Managing Editor Patrick O’Rourke thinks so. The tech is so on level that it blurs the road between actual and pretend, with no apparent markers that scream “AI-generated!” This lack of transparency could make any photograph suspect. Whereas Reimagine has some guardrails, in case you’re intelligent along with your wording, you may skirt them fairly simply. What do you concentrate on Reimagine?

ChatGPT, alternatively, didn’t want the phrase “numerous” to create a picture of enterprise professionals with multiple pores and skin tone. Once more, although, with massive teams of individuals, DALL-E tends to soften faces with generally terrifying outcomes.

DALL-E vs. Grok: Which AI creates the higher photographs?

Grok could be the youthful AI, but it surely produced photographs that had been extra lifelike than the cartoon-like photographs nonetheless created by DALL-E. X’s AI additionally tended to create these generations sooner. The premium subscription to X additionally prices $8, whereas, if you would like the newest model of DALL-E, you may want $20 for the ChatGPT subscription. (Although the DALL-E dataset can also be behind Microsoft Bing’s free AI).

Nonetheless, the less content material restrictions imposed by Grok is not at all times factor. Of the 2 AIs, Grok appeared the extra prone to break copyright and use a licensed character. The flexibility to create people who appear like celebrities additionally provides Grok the better potential for misuse creating deep fakes for political propaganda and pretend information.

DALL-E 3

Offers extra cartoonish, much less lifelike photographs, however probably creates much less of an ethical and authorized conondrum than X AI’s Grok. To entry the newest model, customers must pay $20 for ChatGPT premium.

Grok

Owned by X (previously Twitter), Grok is new and extremely lifelike than OpenAI’s DALL-E, which implies customers could must be extra cautious in relation to authorized implications. A subscription prices $8.

Developer

X (previously Twitter)

Subscription value

$8 for premium