I Tested Three of the Biggest AI Image Generators, and One Came Out on Top (Barely)

We may earn a commission from links on this page.

Google’s Gemini AI app has been topping the “most downloaded” charts on both the Apple App Store and the Android Store ever since the company added a free image generation feature, called “Nano Banana,” back in August. Of course, Google is hardly the only huge tech corporation with an AI assistant that can make images right on your phone.

I wanted to know which mobile AI image generation tools is best, so I pitted three of the biggest—Google’s Gemini (with Nano Banana), the iPhone version of OpenAI’s ChatGPT, and Meta’s Meta AI—against one another in a not-so-old-fashioned image-generation throw down. While there was ultimately a winner, the results weren’t exactly clear cut.

How I tested these tools

I wanted to compare how each app handled the same basic prompts to generate images an average user might want to create.

To test their photo editing acumen, I asked the different models to remove an object from a photo and to expand the background of a photo. To test their utility for straightforward purposes, I asked them to create a cover for a brochure. And to test their “creativity,” I asked them to put a celebrity in a surreal situation, draw a one-panel comic, and make an image of Frankenstein doing stand-up comedy.

Here’s how it went.

Removing an object from an image

For the source image, I used the below photo of my mom, and the prompt “Remove the cup from the subject’s hand.”

Here are the results:

Gemini object removal

ChatGPT object removal

Meta AI object removal

Winner: Gemini

Loser: ChatGPT

While all three tools removed the cup, Gemini added a fairly natural-looking hand held in a jaunty position that suggests my mom has just made a really good point. Other than that, Gemini mostly left the original image alone, just like I asked.

Meta AI’s made the fingers look cartoonish and left the hand in an awkward looking “holding a cup” pose, making the image look like someone did a bad Photoshop.

I am not sure what ChatGPT is doing here. It seems to have removed my mom’s entire right arm instead of just the cup. It smoothed out wrinkles, took out stray hairs, changed the entire color palette to be more orange, and even subtly shifted the direction my mom is looking. I asked for none of this, and all it of it made the photo worse. ChatGPT, You made my mom into an AI-ghoul; you’re doing too much.

Expanding a photo’s background

For the “expand the background” challenge, I used this selfie, and the prompt “Expand the background on this image and remove the sweat stain.”

Gemini background expansion

ChatGPT background expansion

Meta background expansion

Winners: Gemini and ChatGPT

Loser: Meta AI

There are really only two competitors here, as Meta doesn’t do background expansion.

Gemini was more ambitious this time out: It expanded the background further, and did a nice job approximating what the parts of my bike and bike rack it couldn’t “see” actually looked like. It even added a distant car. But it also changed the shape of the mountains behind my head for some reason, and turned down the red-tint—more flattering, maybe, but not requested.

ChatGPT was more modest in its background expansion, and while it didn’t mess around with the color scheme, it did give my skin that weird plastic look common to many AI images.

I consider this one a draw: adequate work from everyone. Except you, Meta AI.

Generating an image for a brochure cover

For this test, I let each tool have more “creativity,” but also provided some clear context and a suggested style, via the prompt “I’m making a brochure for my country club. Generate a painterly image of two rich people playing tennis.”

Gemini brochure cover

ChatGPT brochure cover

Meta AI brochure cover

Winner: ChatGPT

Loser: Meta AI

The winner here is obvious. ChatGPT’s output looks “painterly,” as requested, and the placement of the two figures suggests a friendly game of mixed-doubles.

I found Gemini’s generic depiction of “rich people” to be kind of funny, especially with the mansion in the background, but that is not what a painting looks like, and that is not how anyone plays tennis.

Meta’s depiction of “people playing tennis” isn’t funny. Its result looks like an Exhibit A in a high-profile divorce case, and domestic violence is not a joke.

A famous person in an unlikely situation

To test how each program would handle creating the likeness of an actual person—a dead person, to stay on the safe side—I fed each tool this prompt: “Generate a photo of David Bowie going cave exploring.”

Bowie spelunking by Gemini

Bowie spelunking by ChatGPT

Bowie spelunking by Meta AI

Winner: Meta AI

Loser: ChatGPT

This time, Meta’s the hands-down winner. I asked for a photo of David Bowie, and got something like a photo of David Bowie. I like that Meta chose an older-Bowie, but not ancient-Bowie, as if he’d taken up cave exploring to clear his mind and ponder his future after the commercial failure of 1989’s Tin Machine.

I’m not sure what Gemini is going for here: Bowie with a light saber made out of a crystal and wearing a colander with lights for a hat? Bowie was cool, man.

But ChatGPT is the big loser, for being cowardly and not generating an image at all.

Drawing a one-panel comic

I like asking AI to tell jokes, because I like to see hard evidence that there’s still something people can do better than robots. Expecting AI to actually be funny is as foolish as—I couldn’t come up with a simile, so I asked chatGPT, and it said, “…asking a goldfish to explain quantum physics while juggling flaming marshmallows.” Ha ha ha.

Anyway, I thought if I gave AI guidelines and a model of something funny, maybe it would come up with a good comic. Here’s the prompt I used: “I’m making a one-panel comic in the style of The Far Side. Generate an image for the caption: ‘The real reason Larry was late for work.'”

Here are the results:

The Far Side by Gemini

The Far Side by ChatGPT

The Far Side by Meta AI

Winner: Gary Larson

Loser: Comedy itself

Are any of these comics funny? No. But I think Gemini provided the most interesting result: It kind of made a joke, but it also made me think. If the joke is that Larry was late because it was his goose’s birthday, why is there a hole in the door? Why is the goose so mad? Why is there a suitcase full of money and a UFO? Sometimes I didn’t understand The Far Side either. I also appreciate that Gemini didn’t copy Gary Larson’s drawing style at all, but did add the signature “Gary Larnson.”

Meta AI’s comic is just lazy. I’m not convinced it’s even reading my prompts.

ChatGPT’s result looks the most like The Far Side, without being a direct copy, and the signature is even spelled correctly. But it doesn’t capture any of the weird spirit of the source material. In the end, it’s way more obvious and workmanlike than Gemini’s left-field approach.

And it’s also worth noting that here I ran into one of the main limitations with ChatGPT’s iPhone app when compared to Meta AI and Gemini: I ran out of tokens for the day and had to wait 24 hours to make the image. Output quality aside, if you’re interested in iterating and improving your result, or you just want to make a ton of pictures, five a day on the free tier will certainly hamper your, uh, creativity. Your solution is to upgrade to the paid version for $19.99 per month.

Frankenstein doing stand-up comedy

I next asked these programs to generate images of Frankenstein doing stand-up comedy, because that’s the kind of person I am. The prompt: “Generate a photo-realistic image of Frankenstein doing stand-up comedy.”

Here are the results:

Frankenstein doing stand-up comedy by Gemini

Frankenstein doing stand-up comedy by Chat GPT

Frankenstein doing stand-up Comedy by Meta AI

Winner: Everyone!

I can’t choose a favorite here. ChatGPT followed the prompt most closely, and depicted an expressive Frankenstein having a good night.

Gemini went way off-script, but sometimes you don’t know exactly what you want, and it turns out I wanted a crowd made up of both people and Draculas, with a monster with a lost expression, like he’s trapped between two worlds.

Meta AI’s miserable monster seems to be saying “We belong dead!” which I appreciate as well. So it’s a three-way tie.

Note: Not one AI pointed out that “Frankenstein” is the name of the doctor, not the monster.

The ultimate test: Recursive image generation

Every blog post needs an image to accompany it, so as a final, ultimate test, I fed this entire article into Gemini, ChatGPT, and Meta AI with the prompt: “Generate an image to accompany this blog post.”

Gemini recursive test

ChatGPT recursive test

Meta AI recursive test

Winner: Gemini

Loser: Art

Meta AI seemed determined to covertly compare tennis to domestic violence, and ChatGPT’s grid approach is staid, but I gotta hand it to Gemini for at least understanding the assignment.

(The real test is whether Lifehacker’s editors have left the image in place at the top of this page or sent me a terse message saying, “Steve, take that garbage down immediately.”)

Overall winner: Gemini (but not by much)

There’s a reason everyone has been downloading Gemini to mess around with Nano Banana—it’s really good. It’s not perfect—in my tests, ChatGPT’s image generation engine was better at generating different styles of art from scratch—but Gemini can whip up pictures fast that are often surprisingly close to what you want.

And Gemini is free, whereas ChatGPT’s app costs $19.99 a month for unlimited pictures. Meta AI is also free, and its results have a goofy charm, but it fails to properly understand prompts more often than the other two models, and doesn’t have some useful functions, like expanding backgrounds. (It did a good job with Bowie, though, I must admit.)

Original Source: https://lifehacker.com/tech/which-ai-image-generators-is-better-chatgpt-gemini-or-meta-ai?utm_medium=RSS

Original Source: https://lifehacker.com/tech/which-ai-image-generators-is-better-chatgpt-gemini-or-meta-ai?utm_medium=RSS