Exploring Midjourney's Prowess in Generating AI-Powered Artworks: A Comparative Analysis of Models

Exploring Midjourney’s Prowess in Generating AI-Powered Artworks: A Comparative Analysis of Models

Quick Links

arguably the easiest AI art generation tool both in terms of using it and in terms of getting amazing results, but it’s easy to forget it wasn’t all that impressive at launch. Lets see how far it’s come!

Our Methodology

Miey works by accepting text “prompts” describing what sort of image you want it to generate via the Discord app. If you want to learn how to use the software, there are many awesome Miey guides on the web that will have you up and running in no time. For this article, we’re focused on comparing how much Miey’s technology has improved in the 15-odd months between V1 and V5.1, which is the latest as of this writing.

Related: What Can You Do With AI Generated Art? ey allows you to manually switch between any of its model versions using

    `/settings`

which will give you these options.

Here we can select between the different Miey versions. Incidentally, the “Niji” options are models trained specifically to generate images in the Japanese Anime and Manga style, so if that’s your thing give them a try.

All of the images generated for this article will use the High Quality setting, with the default style level, which determines how strict or loose the model behaves when following the prompt.

We’ll be using the same five prompts with every model so that you can make a direct comparison. We won’t be using any special Miey parameters and will stick to plain text prompts in plain English.

We used ChatGPT-4 to come up with five prompts to test Miey, representing a variety of challenges. These are the prompts we’re using:

Oil painting of a surreal landscape
Photo of a busy downtown street in the 1920s
Concept art of a futuristic cityscape at night
Painting of a white tiger in the snow
Portrait of a surprised woman

Miey provides four images in response to each prompt. We’ll pick the one which (in our opinion) is the best of the four to present here for comparison. So let’s see how much it’s improved in the short time it’s been around.

Midjourney V1 (February 2022)

We’re firmly in the realm of dream-like nightmare fuel with V1 of Miey. At a quick glance these images give you the impression of what they are, but as soon as you actually look at them it all breaks down. Human faces are particularly disturbing.

Miey V2 (April 12, 2022)

V2 doesn’t seem like a huge leap over V1, but there’s a definite improvement in composition, especially noticeable in the 1920s street scene. The white tiger definitely has a better outline than before, and the woman’s face is ever so slightly less creepy.

Notice how, in some images, there’s a sort of pseudo-signature on the piece. This is because the AI has learned that images of certain styles have signatures, and so it tried to reproduce something that looks like one.

Miey V3 (July 25, 2022)

With V3 we’re definitely seeing a significant improvement in coherence and composition. The details are clearly off without too much scrutiny, but the amount of random noise is definitely reduced and the images make more sense to the human eye.

Miey V4 (November 5, 2022)

Now we’re talking! V4 shows a dramatic improvement in composition, details, and how dynamic and imaginative the images are. Just over half a year since V1, and these images are definitely in the realm of usability, often just as they are without further editing. It takes significantly more scrutiny to pick out tell-tale AI flaws.

Miey V5 (March 15, 2023)

The improvements with V5 are a little more subtle than before, but they are substantial. Apart from the technical improvement in detail and composition, these pieces seem more “imaginative”, with even more dynamics, broader choices of colors, and overall much more pleasing presentation. Although V4 certainly moved the resulting images into the grey zone between the best human technical proficiency and what AI generation can do, V5 makes it virtually impossible to tell that these images are AI-made when Mifiring on all cylinders.

V5 is also notable for addressing one of the biggest problems in AI image generation: hands!

MidJourney/Sydney Butler

The above image from V5 shows that the issue has by and large been solved, although here you can still see that one of the hands might either have a missing pinky, or its simply obscured. Edge cases like these still slip through, but it seems the days of people with 7-fingered hands might be almost over.

Miey V5.1 (May 3, 2023)

While the last model update we’re looking at here may only be a “.1” update, there’s certainly no mistaking the improvement. Building on the fundamentals of V5, V5.1 polishes up the minor flaws in V5 while, to our eye, showing an unmistakeable improvement in expression and overall artistic flair .

It’s hard to imagine where Miey can go from here, but the amount of progress so far is simply mind-blowing, and we can’t wait to see what’s possible next.

Tech Haven