Photoshop vs. Midjourney vs. DALL-E 3: Which AI image generator wins my 5 tests?

Adobe’s AI announcements this week were pretty significant, especially the AI-based generative vector features in Illustrator. Photoshop has had a slew of small improvements, but nothing as recently game-changing as the generative fill feature announced at the beginning of the year.

That said, Adobe just announced that Photoshop can now generate an image given a blank canvas, putting it in the same category as text-to-image generators like Midjourney and DALL-E 3. To that end, I wondered how Photoshop’s text-to-image feature compares to the others, especially in terms of creativity.

Also: Two ways you can build custom AI assistants with GPT-4o – and one is free!

So that’s what I’m doing in this article. I’m feeding the same prompt to Photoshop, Midjourney, and DALL-E inside ChatGPT using GPT-4o. I’m not adding any of the styling options each tool offers, except in the case of one test where I ask for “dieselpunk style” in the body of the prompt.

Which tool produces the better art? You’ll have to be the judge of that. Let’s get started. To see any image bigger, click the little square icon in the upper right corner.

Test 1: Helicarrier

The first image that came to mind was the helicarrier from Marvel’s comics and movies. This is an aircraft carrier improbably held aloft by turbofan engines. The physics is questionable, but the cool factor is off the rails.

Here’s the prompt:

Aircraft carrier flying in the sky, held up by four upward-facing turbo-propellors in round fan housings, carrying a squadron of fighter jets on its deck

For each of these tests, I’ll show you what I think were the best images from all three AIs side-by-side so you can compare them.

Here’s what we got:

As you can see, Photoshop missed the boat. Literally. Its knowledge base probably doesn’t understand what an aircraft carrier is. If you look closely, the bombs do look like they have little propellors on them, though.

Midjourney generated a good helicarrier-looking craft, but didn’t include any of the turbofans. DALL-E did, but even though I specified “upward-facing,” it clearly didn’t know what to do with them, so it just sprinkled them throughout the image.

Also: The best AI image generators of 2024: Tested and reviewed

Midjourney always creates four images, so for each of the tests, I picked the one I thought was the best and used that for comparison.

I decided to go back to Photoshop and turn on one of its styling options, Steampunk. As you can see, Photoshop produced a much more interesting image. It’s not exactly an aircraft carrier, but it does have a conning tower. And if you look below, that area might be where jets are stored. I’m not sure, but it’s definitely a much more fun image.

Text-to-image success rating:

• Understood overall spirit of prompt: Midjourney and DALL-E
• Included all the details specified: DALL-E

Test 2: Giant dieselpunk robot

Wikipedia defines “dieselpunk” as “a retrofuturistic subgenre of science fiction similar to steampunk or cyberpunk that combines the aesthetics of the diesel-based technology of the interwar period through to the 1950s with retro-futuristic technology and postmodern sensibilities.”

If steampunk is all about the Victorian age with steam engines and gears, the dieselpunk era is 1940s and 1950s with the chunky diesel-fueled technology of the time.

In any case, I fed the AIs this prompt:

A colossal robot standing as the protector of a glorious city, with citizens bustling around and commuting to work, done in dieselpunk style

To be fair, all three AI tools did a really great job:

Photoshop clearly understood the assignment. Midjourney’s image was the closest to what I had in my head when I issued the prompt, but the robot doesn’t look like a guardian. It also doesn’t look like a threat, since people are just milling about.

Also: How to selectively modify a Midjourney image to create an artistic statement

The DALL-E version, with the robot in front of the city, fits in more with the guardian idea. There’s a bit of a question about whether it confused steampunk with dieselpunk because of the steam engines. Steam engines generally went out of service in the 1930s, but they were still running on some lines as late as 1960, which would mean they could coexist in the dieselpunk era. The people are dressed for dieselpunk rather than steampunk, and the buses in front are clearly dieselpunk rather than steampunk.

Text-to-image success rating:

• Understood overall spirit of prompt: All three, but possibly DALL-E a bit more
• Included all the details specified: All three

Test 3: A Kid in King Arthur’s Court

One of the things that happens when you write an article like this is you cast about for ideas for test cases. I was thinking it would be interesting to see how the AIs did with juxtaposed time periods, since they did fairly well with the dieselpunk era in representing the citizens.

Also: How to use ChatGPT to digitize your handwritten notes for free

I happen to be a fan of alternate history and time travel, so as I thought of this, my mind went to a little film from the mid-1990s that dropped a Little League player into medieval times (the era, not the restaurant). That movie was A Kid in King Arthur’s Court. It’s not a classic by any means, but it was a fun movie. And so, that gave us this prompt:

A young teenage boy, wearing a modern Yankees-style baseball uniform with hat and mitt, stands in the bustling center of a medieval court with citizens and knights in armor

Here’s what the AIs produced:

The Photoshop image was the most realistic. It could well have been taken of a real kid on a movie set. Unfortunately (but probably predictably because of Adobe’s licensed-only media library), the kid was not in a Yankees uniform. That said, the AI did a masterful job of including baseball-style cues in the costume it gave the kid.

Also: How Adobe manages AI ethics concerns while fostering creativity

Midjourney created the most cohesive-looking image, although the kid looks as unhappy as a kid might look if he suddenly landed in King Arthur’s era. The other images showed kids looking even worse, as if they’d been in medieval Europe as long as Martin Padway had been in L. Sprague de Camp’s Lest Darkness Fall book series.

DALL-E was the least impressive. It did meet the spec, but the image has a very uncanny valley feel about it that the other two images did not.

Text-to-image success rating:

• Understood overall spirit of prompt: All three
• Included all the details specified: Midjourney and DALL-E

And now that we’re done being inspired by our inner Mark Twain, let’s move on back…to the future.

Test 4: Back to the Future

Another issue I had in constructing tests for this was the need to use references that I could be sure everyone would understand. There are a bunch of great contemporary movies, even in science fiction, but I’m not sure we all have the same frame of reference.

Also: How I used ChatGPT and AI art tools to launch my Etsy business fast

But by channeling these old classics through the AI, I can be pretty sure most of our readers know what to look for and what to expect. That, of course, brings us to Back to the Future. This time, I wanted to give the AI a movie title and see what it did with it. Here’s the prompt:

Create an image that represents the movie “Back to the Future”, include themes of the 1950s, the DeLorean time machine car, and the boy with a skateboard

Here’s what we got:

Photoshop… I mean, what can you say? The kid popping out of the windshield of a car is a little disturbing. Again, the licensing-limited nature of the Adobe Firefly knowledge base held back the image generation, but the AI did try to include something of the 1950s feel. Here are the other two Photoshop images:

Midjourney did an amazing job with all four of its images, although the first one missed the car:

DALL-E put up a fight. Even though, in a previous test, DALL-E was perfectly happy using copyrighted content, this time it just said no. When I fed it the above prompt, DALL-E in ChatGPT responded with:

I was unable to generate the image as requested due to content policy restrictions. Please let me know if you have an alternative prompt or if there are any adjustments you’d like to make.

I asked ChatGPT:

Rewrite the prompt so it reflects as much of the intent as possible without violating policy restrictions

ChatGPT did not give me back a new prompt, but it did give me the image I showed above, which is very close to the intent of the project.

Text-to-image success rating:

• Understood overall spirit of prompt: Midjourney, with DALL-E getting honorable mention
• Included all the details specified: Midjourney, again with DALL-E getting honorable mention

Test 5: ZDNET’s motto

ZDNET’s motto is “Tomorrow belongs to those who embrace it today.” You see it at the top of every page:

I wondered what the AIs would do with it, and so I fed that phrase to each of them. Here’s what came back out:

So, Photoshop bodars the emieragge tiday, you know? What is there to say to that? Midjourney, on the other hand, created a very interesting and pretty image. It does have the feel of ZDNET’s motto, even if the image itself is hard to describe. And, despite a very minor lettering issue, DALL-E did a fine job of representing the future and its possibilities.

Text-to-image success rating:

• Understood overall spirit of prompt: Midjourney and DALL-E
• Included all the details specified: Midjourney and DALL-E both produced an image with the feelings the motto conveys
• Went to a dark place nobody understands: Photoshop

Which AI wins?

Unfortunately, it’s not Photoshop. Photoshop mostly succeeded on two of the five tests, but pretty much failed the others. Midjourney was the most consistently successful, but DALL-E came pretty close.

Also: I asked DALL-E 3 to create a portrait of every US state, and the results were gloriously strange

The thing about Photoshop is that while Photoshop’s text-to-image isn’t all that great, Photoshop has the very best generative fill capability of any of the AIs. I’m still very likely to create images in Midjourney (and sometimes DALL-E), and then bring them into Photoshop for a generative fill touch-up.

What do you think? Which of these tools do you use? Which AI do you think did the best in our tests? 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, and on YouTube at YouTube.com/DavidGewirtzTV.

Source Link