Can Midjourney AI outperform a human artist? We tried it out at work.

These days, generating realistic images using text is just a part of reality. So, should human designers start to delegate (at least some of) their tasks to AI? Alena Kirdina, Senior Product Designer at Evil Martians, decided to try it at work and find out.

Why Midjourney AI?

For the experiment we’ll cover here, we’ll go with Midjourney AI. However, there are several other notable AI models for text-to-image are worth mentioning: like the OpenAI backed DALL-E, and Stable Diffusion.

But, in our view, Midjourney offers the coolest creative options. Beginning in November 2022 with version 4's alpha, it began producing images with amazingly high quality, and it’s almost impossible to distinguish these results from human-created illustrations.

Introducing the experiment

It wasn’t too difficult to track down a suitable work task to pit humans against machines. Nearly every week, on our blog at Evil Martians, we publish a new article. And we pair each article with an original, human-created illustration that fits the theme of the article. We work with real illustrators to produce these images, and we usually communicate what we’re looking for via text. Midjourney AI understands text, too. Bingo.

We’ve worked with our familiar roster of talented, reliable freelancers for quite a while, so the amount of art direction we need to provide them is minimal. Despite this, the turnaround time to get an illustration just right can take around one or two weeks.

Our images also require a particular style: the vast majority of the images have an isometric 3D view, with a “plastic”, vivid apperance, and solid backgrounds with dimensions usually set at 1200×1000. We’ll use these values as technical parameters.

An example of an image in our standard style

Defining experiment metrics

We’ll need to define some basic parameters in order to properly assess the results:

Overall creation time. This includes everything from the initial request to the end result.

Active art direction time. We’ll calculate how much time was needed to refine the composition with our human illustrator versus prompt adjustments with Midjourney.

Cost. We’ll compare the approxmimate cost of both approaches.

Finally, which picture looks better?

The cat’s out of the bag

Let’s get started. Here’s (roughly) the prompt we sent the human illustrator:

“We’re writing an article about migrating from CircleCI to GitHub Actions. Their logos are a circle and a cat (see attachments). So this is the idea — the cat is playing with a frisbee, ball, or other round object. Can you design the cutest kitty in the world? You can even make it a little Martian if you want.”

Four days later, we got back some detailed renders in a lovely “toy” style, which met all technical requirements:

The human artist’s first draft

We went with the green version. We did have a small request: change the shape of the cat’s tail and the muzzle size. The adjusted image came the following day, and we were all super happy with it.

The final result after some tweaks

In the end, the entire process required six days, and half an hour, max, of active direction.

For our AI illustrator, Alena needed to translate our plain English prompt into a text prompt to send via Discord:

/imagine: pixar style 3d image of the cat in a spacesuit playing with a ball on solid green background --ar 3:2 --v 4

The last two setting commands are as follows: ar is aspect ratio, and v 4will specify we want to use the latest Midjourney release to get the highest quality results possible. pixar style 3d defines the style that I’d like the AI to reference.

Here are the results, which were immediately sent back:

This was pretty far away from what we were after

Alena thought it looked amusing, but the “plastic” style from the Evil Martians’ blog is missing. She decided to add “plastic”: pixar style 3d of a toy plastic cat in a spacesuit playing with a ball on solid green background --ar 3:2 --v 4

Some of these look pretty strange

The new results were certainly “plastic”, but slightly frightening, too, with some appearing to have extra paws.

Again, Alena repeated the prompt, trying to turn the cat black and more like GitHub’s mascot: pixar style 3d of a black toy plastic cat in a spacesuit playing with a ball on solid green background --ar 3:2 --v 4

Where’s the action?

She repeated this several times, but it seemed to always lack dynamics. Thus, she tried to add some text that would indicate motion: plastic toy cat in a spacesuit jumping and playing with a ball, black on solid green background, Pixar style isometric 3d render --ar 3:2 --v 4

Go kitties!

Yes! Now it appears quite dynamic, and much more interesting. Adding the term “jumping” instantly gave the results the missing X factor. From here, some variations were generated:

A star is born

From here, Alena upscaled a winning selection.

Since Midjourney isn’t so good at working with the solid, uniform backgrounds required by our blog, we needed a tool to smooth things out. Luckily, the AI-powered Remove.bg can do just that.

And there it is! All in all, it took around an hour and a half of focused work.

So, who won?

For this first experiment, the human illustrator certainly required less art direction, while the AI delivered a final working result days faster. In terms of cost, the AI was also vastly cheaper. As for the quality of the final result, that’s a bit subjective! (For the record, we used the human-made illustration for that article.) What do you think?

And this was just round one! If you’re interested in seeing more battles, head to the full article on our blog: we ran the same experiment multiple times, and you’ll also be able to see more detailed technical breakdowns and cost comparisons there. Check it out!