The best AI image generators of 2025: I retested the top tools – there’s one clear winner

Follow ZDNET: Add us as a preferred source on Google.

ZDNET’s key takeaways

• Google’s Nano Banana Pro earned a near-perfect score.
• ChatGPT image ranked second; others often mangled text and faces.
• Nine tough prompts reveal which AIs are worth subscribing to.

When generative AI first hit the scene in early 2023, we were amazed by two capabilities we hadn’t seen before: the ability for the AIs to write comprehensive content, and the ability for some AIs to generate graphics and photo-realistic images.

Clearly, AI models have come a long way in just three years. In my recent review of the best chatbots, I also listed the best free AI tools for generating images. In this article, I’ll take you a little deeper into the paid tiers of AI image generation tools.

Also: The best AI chatbots of 2025: I tested ChatGPT, Copilot, and others to find the top tools now

My test suite falls into four main categories: revising existing images, generating original images, adding text to images, and leveraging pop culture. For convenience, I generated all the images across all the tests in 1:1 square ratio, mostly so they could be displayed side-by-side easily.   

I generated over 50 images using six popular AI image generators. Each image generator was subjected to 30 individual evaluations, and from that came an aggregate score.

We have a winner

There’s a winner — a no-contest, total knockout winner. This is a first in my AI evaluations. Usually, there are tools that each do a fairly good job. However, in this set of tests, there’s an absolutely clear standout champion: Google’s Nano Banana Pro, part of Gemini 3.

It aced every one of my tests with a near-perfect score of 93%. ChatGPT came in second place, at 74%. The other four ranked between 43% and 54%.

Also: Stop using ChatGPT for everything: I use these AI models for research, coding, and more

All six contenders did a fine job when asked to create a logo. Beyond that, performance was all over the map.

The bottom line is, if you want serious AI image generation, spend $20/month for Gemini 3, and you’ll get Nano Banana Pro. That’s it. There is no other possible conclusion.

Although some tools offer additional features, the primary difference between the paid and free versions is the amount of usage you get. For example, Nano Banana Pro cut me off after two images before I upgraded.

Also: How to install and configure Claude Code, step by step

And, just for the record, I bought all my own subscriptions to do these tests. Nobody is paying for this conclusion. I’m as shocked about the Gemini conclusion as many of you might be.

OK, so let’s get started.

Google Gemini 3 Nano Banana Pro

Best AI image generator overall (by far)

• Overall score: 93%
• Cost: $19.99/mo as part of Google AI Pro

AI lets you do photo recontextualization in minutes. Basically, you can feed an AI an existing image and ask for it to be modified into something completely different. The formal test is feeding a picture of me out on a walk and asking the AI to put me in a US Navy admiral’s uniform on the bridge of an aircraft carrier. I know. You have your fantasies; I have mine.

So, here’s my starting image.

And here’s the recontextualized image (below). Notice how it kept my face and glasses, but even changed the angle I was facing. It put me properly on the bridge of a ship and added binoculars to my hands. I mean, I make that look good. 

OK, to be fair, there are issues with Navy personnel (admirals included) and beards. But I can live out my AI fantasy for the duration of this article, right?

Next, I wanted to see how well it would clean up old photos. I found a photo of me standing on the USS Ling when I was a kid. Yes, it was my Bob Ross hair phase. The problem was that the image was too dark to see. On the left is the original. On the right is the Nano Banana cleanup.

Where this got interesting was my next test. I fed it an image of a truck and told it to clean up and colorize the image. Instead, it first gave me the same Bob Ross hair picture of me above, but colorized. When I once again asked it to do the truck, it put me in the admiral uniform in the image. 

It’s quite shocking to see side-by-side pictures of myself, separated by decades. While this was a cool image and gave me food for thought, it wasn’t what I asked Nano Banana Pro to do. Points were deducted.

There’s one thing I don’t like about Nano Banana Pro, and that’s that it puts a watermark in the bottom corner of each image. You could use another AI to clear out that watermark, but it’s annoying.

So, what happened with the truck? I started a new session and then uploaded this image of the truck. I think my dad took this image 40 or 50 years ago.

Here’s how Nano Banana Pro cleaned it up (below). Notice the lettering on the truck is mostly accurate and readable. However, my editor noticed that it really says “BADICLOGICAL DEFEKSE,” costing Nano Banana Pro another few points. 

Text is definitely a problem for most AIs. Also, I’m fairly sure that truck was mostly blue, but Nano Banana couldn’t have known that, and it colored it like many other emergency vehicles of the time.

Now, let’s look at creative image generation. On the left is a logo for Space Coast Studios, which I described as a video production studio. On the right, I asked for a medieval librarian in a candlelit stone library.

As you can see, Nano Banana picked up all the themes for the logo. Space Coast is in Florida, so there’s a palm tree. There’s a rocket, film, and good text. On the right, we have our librarian, lots of candles, candlelight, and stone walls.

Next up is a task I expect many people will be using AI image generators for: social media. I gave the AI two related assignments to create images for Facebook. The first was to generate a senior adult holding a phone, along with some text. 

The second was to generate a student in a cozy environment, specifically with a MacBook Pro and a cup of coffee. Gender was not specified in either assignment.

I want to point out that the text is done properly, fits over the image cleanly, and there are no errors. But I’m more amused that the guy is holding what’s clearly an iPhone. Yep, Google’s AI generated an iPhone when asked to show a “flagship smartphone.”

Finally, I ran a couple of pop culture-inspired prompts. These test how the AIs deal with cultural references, whether they are copyrighted intellectual property or ideas inspired by the culture.

Also: How I use AI to bring my kid’s art to life

For the first image, I asked for a poster for a fictional “fourth Back to the Future movie,” where Marty McFly goes back to 1920s New York.

The second pop culture image I requested was “Nightmare Before Christmas, Tim Burton style,” where I wanted an IT professional in a data center.

Nano Banana Pro did a fine job with the Nightmare style. But take a look at the details in the Back to the Future piece. It got the DeLorean and the skateboard. It did a great job with the text, not only evoking the “Back to the Future” logo style, but putting the text “New York City 1925” on a marquee just like they looked in New York back in the day.

It added the tagline, “The Roaring Twenties Got Heavy,” referencing how the ’20s were described and referencing one of Marty’s favorite terms. It also put Marty into period clothing.

Just wow. Across all my tests, other than that one time when I had to start a new session, and the minor truck text issue, Nano Banana Pro didn’t get anything wrong. It gets a nearly perfect score.

ChatGPT

Best for natural-language iteration in chat

• Overall score: 74%
• Cost: $20/mo as part of ChatGPT Plus

Back in the early days of generative AI, we discussed ChatGPT as the text-based chatbot and DALL-E 3 as OpenAI’s text-to-image AI tool. DALL-E, as a standalone tool, has pretty much fallen out of use in favor of the image generator inside of ChatGPT.

Just before writing this, I reached out to OpenAI and asked them what that tool is called, and I think this explains why ChatGPT is 19 points behind Nano Banana Pro. I was told that ChatGPT’s image generator is “GPT-4o image generation in ChatGPT.” 

Also: Is that an AI image? 6 telltale signs it’s a fake – and my favorite free detectors

I originally ran this set of tests against that image model. But then, OpenAI introduced its new ChatGPT Images update, which, in some ways, is a vast improvement over the previous version. For my first look (and a lot of fun with image transformations), take a look at this article: I tested the new ChatGPT Images – it’s a stunning improvement, and enormously fun.

Since we do have a new model, I redid the ChatGPT tests, and I’m using the new one in these ratings. 

Let’s start with the photos. ChatGPT lost a bunch of points here. It’s fair to say, photo retouching is not a strength of this model.

The admiral recontextualization is pretty good. That looks like my face. However, the uniform is somewhat problematic. It’s not clear what rank insignia are being used, and they seem to be a weird cross between those of a Lieutenant and an O-8 Admiral. 

The kid picture is good. In the earlier ChatGPT image, ChatGPT put a random adult’s face on the kid’s body. Here, it’s the original face. 

Also: I’ve been testing AI content detectors for years – these are your best options

ChatGPT did a nice job on the truck cleanup, except the lettering changed from what should say “RADIOLOGICAL DEFENSE” to “RADIO CHEMICAL DEFENSE,” with an extra space between CHEMICAL and DEFENSE. Some points were lost, but not nearly as many as in the previous version.

As for the logo, ChatGPT did fine, although I would have preferred some reference to Florida, palm trees, or sun. The librarian was pretty much what might have been expected from the provided prompt. No points were lost.

ChatGPT did almost fine with the two social media posts. I noticed that although the text is correct, it placed both blocks of text in exactly the same place. Moving the text up on the picture of the man would probably have been a better design choice, but it’s good enough. 

I did take a point each off because the lighter font style and poor contrast make the text a little more difficult to read.

That brings us to our pop culture references. This is where things get a little odd. 

In my first look at the new ChatGPT Images update, I basically recreated a scene from a classic Star Trek episode. Not only did ChatGPT produce wonderful images, but it also effectively duplicated the alien creature right out of the episode, all the way down to the creature’s man-in-suit look.

But when I asked for a Back to the Future-inspired image and one inspired by Nightmare Before Christmas, I got this:

That lost the new ChatGPT image generator a bunch of points. In fact, it pushed the new image generator’s ranking from 81% for the previous version down to 74% for this new version.

If you’re curious what we lost, here are the images created by the previous version. The “Back to the Future” image is fine. It’s not as inspired as the one from Nano Banana Pro, but no points were lost. The text logo is good, the Marty character hits all the notes, there’s a recognizable DeLorean, and the city could be New York in the 1920s.

Then there’s our “Nightmare Before Christmas”-inspired IT guy. He does have the gray pallor that most IT folks tend to have. But although I think I can speak for all of us in saying that a third hand would probably come in mighty handy, I would have counted it as an error.

For now, my best conclusion is that pop-culture references in ChatGPT are going to be hit or miss. You might get a Star Trek-inspired Gorn fight, or you might get scolded by the AI and sent to your room without dinner. There’s no predicting.

Midjourney

Best for hyper-creative cinematic imagery

• Overall score: 57%
• Cost: $10/mo for Midjourney Basic

Midjourney was the first dedicated AI service I ever bought, way back in 2023. I used it to help me produce my album launch imagery, a bunch of images for my little test Etsy shop, and I regularly use it to create promo social media posts for my wife’s e-commerce business.

When it comes to originating creative imagery, Midjourney is pretty great. When it comes to image cleanup, recontextualization, or text, Midjourney is… not so much.

As you can see, the admiral image does reproduce a yellow shirt and something of the background from my original picture, but that guy isn’t me, and there’s a ship at the back of the park for some reason.

The kid, as rendered, isn’t younger me, either. Also, Midjourney didn’t clean up the photo. In fact, it made it worse.

Also: Is ChatGPT Plus still worth $20? How it compares to the Free and Pro plans

And there’s a truck, not the truck we started with, but a truck. And for some reason, they stuck a young Kate Mulgrew-looking woman (the Captain from “Star Trek Voyager”) in the image.

All three were provided in 1:1 format as requested, but the images were just slapped onto the 1024×1024 canvas.

Next up are the logo and the library image. Midjourney does provide the logo text correctly. It adds a palm tree and some towers that could look like space. It’s evocative of the topic, but it would have been nice to have a rocket.

Then there’s the library. Here, you can start to see where Midjourney shines. This is almost exactly what I pictured when asking for this image. No points were lost with either of these images.

The social media posts also show some of Midjourney’s strengths. The man looks fine. Even his hands and facial lines seem to go together. He looks real enough. The girl is composed perfectly, except the laptop seems a bit low on her legs (and it’s an older-generation MacBook Pro).

But do you see what’s missing? No text. Midjourney didn’t even try to render the text messages. Points were removed.

Finally, we get to the pop culture references. The “Back to the Future” image lost a few points because there’s no text, and because the city looks way too neon to be 1920s New York. While the Marty in this picture is no Michael J. Fox, I’m not removing points for that. It got the overall look right.

As for the Nightmare IT image, I love this image. This is another example of why Midjourney has such a strong following. When it works, it really works.

Adobe Firefly

Best for commercial-safe images

• Overall score: 54%
• Cost: $9.99/mo as part of Firefly Standard (and included in a bunch of Adobe plans)

Next, we’ll take a look at Adobe Firefly. This is the AI engine inside of Photoshop and other tools, but it also works as a standalone web interface. Firefly (and Photoshop) now let you choose other AI engines (so you can run Nano Banana inside Firefly, for example). That said, all my testing was done using the Firefly Image 5 (preview) model.

Firefly is, by far, the most uncooperative and fussy of the image generators from an overprotective copyright point of view. For example, when I tried to upload an actual picture of myself, I was wearing a Star Trek-inspired Gorn T-shirt. It refused to upload it.

I had to take it into Nano Banana Pro, have Nano Banana remove the T-shirt’s graphic, and then upload it.

The admiral image is barely an admiral’s uniform, and the face isn’t mine. The transformation of the kid image is good, and it’s easy to see. The truck colorization looks great, but Firefly didn’t bother to clean up the image artifacts, which was the main object of the assignment. Points were lost.

As with all the other AIs, the logo turned out to be quite usable. The library image evokes some of the feel I was going for, but if you look at the face, there’s some serious uncanny valley going on. It looks like an AI face generated by 2021 technology. Points were lost here, too.

I had to fight with Firefly to create the two social media posts. Both prompts gave me this error message.

The complaint was, “Including artist names in your prompt does not meet our User guidelines.” Notice that there is no artist name in that prompt. The gotcha was that I used a phrase in the form “[WORD]-style” and that flagged it. When I removed “Facebook-style” from the prompt, Firefly generated the images.

Even though the face of the older gentleman was blurred, that was a valid artistic choice. Opening up the aperture to provide clarity to the foreground works in this context. However, it got the spelling wrong.

For the student image, there are a lot of subtle not-quite-rights. There’s something wrong with his shirt, his arm looks unreal, his hands don’t look quite right, and so on. It’s not quite as bad as the librarian’s face, but it’s not good. Points were lost.

More points were lost for the pop-culture images, which Firefly simply refused to do.

Leonardo AI

Best for fantasy art

• Overall score: 52%
• Cost: $10/mo, but also part of Canva Business plan

Leonardo.AI completely bombed when it came to testing already-existing images. It just refused to come out and play. So, let’s move on to some creative work.

The logo, like all the others, is workable. I personally prefer those that also call back to the Space Coast theme, with sun and palm trees along with the rocket, but this is a perfectly fine logo. As for the librarian, I specified neither gender nor age. So this does meet the requirements.

As for the social media posts, the man seems OK, except for his left inner elbow, which looks a little odd. The woman’s leg seems to wrap around the laptop in an odd way, and she is balancing a cup of coffee on her leg. There’s something just not right about that.

Finally, we get to the last two images from Leonardo.

In my “Back to the Future” prompt, I specifically stated, “Create a poster.” Leonardo is the first AI to actually format a poster. On the other hand, I don’t recall seeing “Back to the Future” four through 10, even if “Back to the Future 11” seems fun, especially with its theme of “HAMY JOB OARY OF DOSE.” Additionally, those cars and buildings are decidedly not 1920s.

We wrap up our time with Leonardo with one of my all-time favorite AI-generated images. Somehow, Leonardo captured the Nightmare feel along with the data center look, but used a photo-realistic Lurch-like character. I absolutely love this image.

Canva

Best to generate art right inside your marketing designs

• Overall score: 43%
• Cost: $15/mo for Pro or $20/mo for Business (includes Leonardo)

Canva is best known for its enormous popularity as a marketing image generation tool. If you want to make a flyer, poster, or social media post, Canva is your friend. When Canva acquired the Affinity Pro image products and released them for free, I became a convert.

Also: I love Photoshop, but Canva’s free Affinity tools won me over (and saved me money)

Along with releasing the Affinity tools, Canva also introduced a whole line of generative AI features, including its own image generator. To be clear, Canva’s internal AI image generator is a separate beast from Leonardo’s image generator, but Canva owns both.

Because Leonardo is $10/month on its own but only $5 more with Canva’s Business plan, I signed up for Canva’s Business plan for this test. I’m probably going to downgrade back to the $15 Pro plan because I don’t find Leonardo all that compelling compared to other tools I’m also paying for.

As for Canva, well, you gotta see this. Let’s look at the photo cleanup and recontextualization tests first, because…there are no words.

I’m not sure, but I think there are trees growing on the aircraft carrier. It did identify the beard, which is more than I can say for the missing truck in what was supposed to be the truck image. And why there’s a disembodied woman’s head floating above the wreckage of something that’s a cross between the USS Ling and Nikola Tesla’s Wardenclyffe Tower is a secret only Canva’s AI brain can possibly understand.

So, yeah. Points were lost. Lulz were had.

Next up is the logo and the librarian. Both are fine. Frankly, if Canva couldn’t create a logo, it would be time to hang up its AI. But it did a good job. It’s the only test category where Canva got full points. 

On the other hand, the social media posts were a disappointment. The man had too much space between “simple” and “powerful,” and the woman’s image has text that lost its way completely. To be fair, Canva has its own specialized social media editing tools, so perhaps the idea is to generate the AI image and then use other Canva tools for text.

Even so, there’s yet another cup of coffee balanced on a leg. Repeat after me: points were lost.

Finally, we get to the pop culture images…or rather, image. Canva made an image for a fictional “Back to the Future 4” movie that did include a Marty McFly character and the DeLorean, as well as the logo in proper BttF format. 

But what’s “WYGLA TLETZ NEW ADIENS?” And who’s the girl? And what’s that thing floating in the sky? And what are those generator-like things on the side of the road? And why is half the road on fire? It’s clearly not back in time to the 1920s.

Look at the above image. Note that there’s a full “Back to the Future” logo there, with the name and even the appropriate logo type. So ask yourself why Canva denied me a “‘Nightmare Before Christmas’ style, Tim Burton style” IT professional?

Where’s the consistency? Why allow one but deny the other? And why so blatantly allow the first, but not even make a try on the other? Points were lost.

Testing methodology

I fed the same exact 9 prompts to each of the six AI image generators above. Here are the prompts and photos used as source material, for your reference:

I created a spreadsheet where I recorded the point scores for each of the images and their prompts.

Admiral photo recontextualization: Points were applied for proper background, proper new clothing, and whether the AI kept the original subject face and clothing.

Dress this man in a US Navy admiral's uniform and place him on the bridge of an aircraft carrier. Do not change the face. Aspect ratio 1:1

Restore black and white image: Points were given if the AI kept the original image elements, kept the original face, body, and clothing, and if the AI properly cleaned up the image.

Restore this old black and white photograph to look like a modern, high-quality studio image. Repair scratches, dust, stains, and cracks. Improve sharpness and clarity without over-smoothing. Enhance contrast, details, and lighting naturally. Preserve original facial features, textures, and period details. Avoid too many dark areas and shadows, make details clear and easy to see. Keep it strictly black, white, and gray with realistic film grain. Aspect ratio 1:1

Restore and colorize black and white image: Points were given if the AI kept the original image elements, properly cleaned up the image, kept the original text on the truck, and converted to color appropriately.

Restore this old black and white photograph and convert it into realistic color. Remove scratches, dust, stains, and damage. Sharpen facial features and fine details while keeping a natural look. Add historically accurate, natural skin tones, clothing colors, and background colors. Match modern lighting and color balance while preserving the original photo's composition and mood. Aspect ratio 1:1

Create a logo: Points were given if the created logo was accurate to the theme, if the style was appropriate, and for text generation accuracy.

Create a retro-futuristic logo for a video studio named "Space Coast Studios" Aspect ratio 1:1

Fantasy librarian: This was a fairly open-ended test, with points given for the quality of the generated image, whether the AI generated an image with candlelight, stone, and a library, and the quality of the generated character.

Create a fantasy scene of a medieval librarian in a candlelit stone library. Aspect ratio 1:1

Studio portrait of a senior adult: Points were given if the AI included glasses, created a realistic face and pose, had the character hold a smartphone, and generated the specified text.

Create a photorealistic studio portrait of a stylish senior adult wearing modern glasses, soft professional lighting, neutral gradient background, holding and using a flagship smartphone. Add short Facebook-style ad copy on the image: Simple. Powerful. Made for everyday life. Aspect ratio 1:1

Candid portrait of a student: Points were given if there was a MacBook Pro (or Air) in the image, the student had a realistic face and pose in the context of the requirements, if there was a cup of coffee (and it wasn’t balanced stupidly on a leg), and if the AI generated the specified text.

Create a natural candid portrait of a student sitting near a window using a MacBook Pro, next to a cup of coffee, soft daylight, cozy home background, relaxed and authentic mood. Add short Facebook-style lifestyle copy on the image: Moments that matter, always within reach. Aspect ratio 1:1

“Back to the Future” poster: Points were given for a realistic Marty, if the AI generated a DeLorean, if the scene was in 1920s New York, and for the generated text and style.

Create a poster for the fourth Back to the Future movie, where Marty takes the DeLorean and his skateboard to 1920s New York City Aspect ratio 1:1

“Nightmare Before Christmas”-style IT person: Points were given if the AI generated a recognizable IT environment, for quality character generation, and for the overall Tim Burton feel.

Nightmare before Christmas style, Tim Burton style, put an IT professional in datacenter Aspect ratio 1:1

Imagine that

So there we are. As you can see, of the six image generators tested, Nano Banana Pro stood out above all the rest. But there are some areas where the others shine as well.

What about you? Have you tried any of these AI image generators yet, and if so, which one impressed you most? Does Nano Banana Pro’s perfect score match your own experience, or do you think another tool deserves the crown? Which features matter most to you: photo cleanup, text rendering, pop-culture prompts, or creative style? And are you paying for any of these tools right now, or sticking with free tiers? Let us know in the comments below.

You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.