Does MidJourney handle consistent character generation? We test it

This material was created in both written and video versions. Watch the video on our Beyond AI channel or continue reading. Thanks!

Watch this material on YouTube:

Introduction

Hi, I'm Ziemek, and today I’m sharing my first impressions of the new feature in Midjourney that ensures consistency of generated characters. I’ll talk about what works, what doesn't, and share my conclusions.

What is Midjourney?

Midjourney is a generative artificial intelligence tool used to generate images. You simply describe what you want to appear in the photo or generated image, and the AI creates it.

Here are some examples showing what users have created recently:

Consistency of Generated Characters

Until now, there was a problem with generating the same character multiple times—each new image would feature a new, different face. Now, that has changed.

I’ll start by showing an example where it works very well. I asked Midjourney to generate a photo of a warrior's face.

I liked face number 2, so I asked it to upscale it.

Once upscaled, you just click "Open in Browser," copy the URL, and in your subsequent instructions, you can use that address as a reference—a specific key provided in the prompt.

I thought it would be interesting to see how Midjourney would generate an image of this warrior on a bicycle in Warsaw. Therefore, I ran a prompt that said "man riding on bicycle in the heart of Warsaw," and as the parameter --cref, I provided the URL I had just copied from the browser.

We can see that it is indeed the same warrior. Midjourney also correctly placed the same character in a group of friends enjoying time together, probably over a beer, and also on a sailboat.

It's clear that it is a similar character. You can even get this warrior married and generate a photo from his wedding party, or focus on his soft side as he holds a kitten in his hands.

Famous Figures in Midjourney

For the purpose of this episode, I decided to test whether I could transport well-known figures into imaginary situations. I uploaded a photo of Donald Trump, and it turned out this wasn't possible. A celebrity or public figure was detected, and Midjourney told me: "No, we won't be doing that here."

Undeterred, I tried with Angelina Jolie, and here it "worked." Indeed, a photo of Angelina Jolie in a park... well, in one of these photos, the character generated by Midjourney resembles the actress.

I didn't succeed with Donald Trump, so maybe I'll succeed with Donald Tusk?

There is something strange about this photo. It has some characteristic features, but it’s not the politician I had in mind.

I tried many more times later with other characters, and the results were not satisfactory. This was supposed to be Maciej Stuhr and Angelina Jolie drinking bubble tea together in the center of Warsaw.

It resembles neither Maciej Stuhr nor Angelina. It doesn't convince me at all.

And what about just Angelina Jolie drinking bubble tea in Warsaw?

In my assessment: no, it doesn't resemble the actress at all. This Warsaw isn't Warsaw either, so we can see that realistic photo generation of specific people using this tool is not quite possible yet.

Summary

What are the conclusions from these first steps? If you want to transport an existing actor, politician, or yourself, I wouldn't have high expectations yet. However, if you ask it to invent a new character—one that Midjourney authored itself—then placing that generated character model in other contexts and situations works well. In that case, you truly have that consistency.

—

If you want to learn more about the fascinating world of artificial intelligence and its everyday applications, visit our YouTube channel – Beyond AI. It’s your guide to the dynamic world of AI!