Stay up to date with SPARK - the sour power newsletter

How Well Can AI Replicate Art History: Part 3 Hilma af Klint

Together let's explore AI's capabilities of recreating and remastering artistry of the past. I'm going to ask Midjourney to describe to me using it's own AI model what the famous image by Hilma af Klint is, and then re-run that description to see how closely we can replicate the mastery of this artist.

About the AI Tool (Midjourney)

Midjourney has quickly become a preferred tool for AI-generated imagery among artists and creative professionals. Launched just a few years ago, it has rapidly become a vital part of a creator's arsenal. We will be utilizing features such as Midjourney's "Describe" and "Imagine" commands.

To learn more about Midjourney's features used in this article, reference Part 1 and Part 2 of this experiment.

✨ Let's dive in! ✨

The Swan (No. 17), Hilma af Klint, 1915

Replication Test Artist and Work: Hilma af Klint's "The Swan"

In the spirit of International Womens Day, let's look at a famous artist in art history who identified as a female.

Klint's Ai Artwork Description

Our first step is to take our sample image (our valued piece of art history) and ask Midjourney (our trusty AI image generator) to give us what it thinks is an accurate description in plain text of the contents of the image.

We'll rund /describe + upload our image into Midjourney and see the results!

Interestingly enough, Midjourney did use the name Hilma af Klint in one of its descriptions, which with such a simplistic piece I was very surprised to see.

Description 1

A simple circle painted in the style of Jean Delville, with a red background and blue and yellow half circles inside the full circle, black in the center of the circular shape, and white on one side of each color. --ar 63:64

(a note: in midjourney the -- lines preface commands in the prompt. in this case --ar 63:64 is a rendered output of our original image's dimensions. in this case "ar" stands for aspect ratio.)

Description 2

A very simple painting of a circle with three colors in the style of Hilma af Klint, with red and white, yellow and blue on an orange background. --ar 63:64

Description 3

A circular, symmetrical composition with three colors – red, blue and yellow – arranged in the center of an orange background. The black color is used to create two large circles on top of each other, symbolizing unity or balance. in the style of Cuno Amiet. --ar 63:64

Description 4

A simple yet powerful painting in the style of jean millet, with two circles of different colors and sizes in the center of an orange background, representing love and affection. The circle on the left is black, while the one to its right has blue, yellow, white, and red stripes. This minimalistic design conveys emotions through color and form, creating a sense of depth within simplicity. --ar 63:64

Hilma af Klint Inspired Geometric Image Generations

How much value is in the artists name?

The first thing to test is an understanding of how the language model responds to the usage of the artists name in the prompt. In our rendered prompts, Option 2 uses our artists name specifically. We only have one test to see the visual weight of this author, so let's see how close we get to our original when we use our artists name!

Description 2 from Midjourney:

A very simple painting of a circle with three colors in the style of Hilma af Klint, with red and white, yellow and blue on an orange background. --ar 63:64

A grid of four images based on the Midjourney prompt description provided to us featuring Klint. Option 2

What is interesting here is we get some pretty close with the top left and bottom right options. We also have some that are in the same playing field but have even more interesting interpretations of the geometry within the main circle.

One thing this has me curious about is how to ask Midjourney for specific spacing and white space within a canvas. When you view the comparison below, you can see that Klint's version has significantly more white space around the center circle.

For the purpose of this experiment, let's compare side by side the artists work against the most favored Midjourney rendering in the grid of four.

Original Klint (left) to Midjourney's recreation (right) of Description 2

I was trying to pick which one I thought was the best representation, and ultimately landed on the last image. It has the most accurate feeling of the geometry of Klint's piece. I noticed after running the prompt that our describe command did a pretty good job of asking for the specific color palette in the original Klint piece, but it did forget an essential tone in the large black area. We still see the left half of our rendered image evoke the same visual as Klint's though, just with different colors.

My least favorite part about this piece is that there are two split background colors. It made it a hard "final" choice for me, so I think if we adjusted our prompt to be more specific with a request from the AI for a solid red background, we would be good!

I would say we got great texture and geometry with this prompt, and it very closely resembls the style of Klint's original artwork.

What does it render without the artist name?

Having examined the description that highlighted the artist's name, let's explore whether shifting our emphasis to the subject matter, rather than the artist, makes any difference.

Descriptions 1, 3 and 4 can help us decipher the importance of different keywords used within an AI prompt.Even without specific defining nouns, by supplying an appropriate array of adjectives, we can test if this approach yields a more favorable outcome.

Let's take a look! 👀

Description 1 from Midjourney:

A simple circle painted in the style of Jean Delville, with a red background and blue and yellow half circles inside the full circle, black in the center of the circular shape, and white on one side of each color. --ar 63:64

A grid of four images based on the Midjourney prompt description provided to us featuring generic terminology. Option 1

The first thing I noticed when running description 1 is that the tone of the image changed when we omitted our author for reference. We still get red and yellow and black, but we get very bold and saturated versions of these colors. In this midjourney prompt we are also referencing another artist "Jean Delville" who has absolutely no known affiliations with geometric-type work in his portfolio. I'm assuming that we got our textures from this reference.

Original Klint (left) to Midjourney's recreation (right) of Description 1

We can see, again, in our side by side comparison that what I really want most of all is a more desirable amount of white space in the final output. The circle takes center stage in both depictions, but one is much more aggressive and in your face compared to our Midjourney rendered images.

Let's see what we get with our other suggested Midjourney prompts!

Description 3 from Midjourney:

A circular, symmetrical composition with three colors – red, blue and yellow – arranged in the center of an orange background. The black color is used to create two large circles on top of each other, symbolizing unity or balance. in the style of Cuno Amiet. --ar 63:64

A grid of four images based on the Midjourney prompt description provided to us featuring generic terminology. Option 3

The AI generated images from description 3 are our most off-base yet. I think these are very creative adaptations and "evolutions" of a Klint-inspired piece. However we were attempting to recreate a very specific geometric style, and we struggled with the use of this Midjourney prompt to get the same shapes within our circle.

In this prompt we also used a different artist for reference in "Cuno Amiet", who again doesn't have any highly recognizable geometric style works in their portfolio. I believe again this was used to help us render that painted aesthetic over subject matter.

Original Klint (left) to Midjourney's recreation (right) of Description 3

We're seeing the same patterns emerge here, with a lack of "perfect" geometry, a forced geometry in the background, needing more white space, and our color palette being a little too bold.

Let's look at our final suggested prompt.

Description 4 from Midjourney:

A simple yet powerful painting in the style of jean millet, with two circles of different colors and sizes in the center of an orange background, representing love and affection. The circle on the left is black, while the one to its right has blue, yellow, white, and red stripes. This minimalistic design conveys emotions through color and form, creating a sense of depth within simplicity. --ar 63:64

A grid of four images based on the Midjourney prompt description provided to us featuring generic terminology. Option 4

This prompt was our longest prompt, and it tried to describe in very specific detail the shape placement and color placement. What we also see though, is the ask for different emotions and metaphors through the artwork to be conveyed. I do believe that the terms like "representing love and affection" and "conveying emotion through color and form" have swayed the results of this piece from being more accurate to our original.

What was also interesting in this prompt is that we lost our red background, which is one of the most critical pieces of the painting due to it's sheer size and volume in the original. The Midjourney suggested prompt actually specified an incorrect color, and used "orange" over "red".

Original Klint (left) to Midjourney's recreation (right) of Description 4

I do think we still get the same "idea" with how this prompt described the geometric ask, while most seemed pretty far off we can see a resemblance in the first image (top left).

So Which Prompt Did The Best?

After comparing all four prompts, in this case the prompt that used the artist's name seemed to render us the best result. I used another tool within Midjourney, which is the "zoom out" feature, to see if I could take it one step further and get a little bit of that white space I was asking for.

There are two options with zoom out within Midjourney, you can select a distance of 1.5x or 2x of zoom. I chose to go for the larger amount since our rendered image really takes up such a massive amount of space in the canvas.

Original Klint (left) to Midjourney's recreation (right) of Description 2 Zoomed Out

In this option, Midjourney hated the idea of all this white space so much that it actually tried to add more circles to the image instead of just giving us more white space. I feel as though I could have adjusted the amount of zoom to a little less than 2x and got the sizing I wanted (with the custom zoom button) but the forceful amount of extra circles added didn't give me the desired white space it just tried filling the canvas with more junk.

My next experiment will have to be "how to get white space in Midjourney!" For now, let's forget the white space and move on to our final thoughts and experiment.

Can We Replicate History with AI?

One crucial insight from these experiments is that Midjourney indeed possesses knowledge of renowned artists and their masterpieces within its model. It demonstrated precision in identifying the name of the artwork and even provided the author's name in one of the prompts. However, introducing additional terms (and some inaccurate) resulted in an unwanted shift away from the distinct geometrics of our artist.

After three different test experiments, it's clear that by carefully crafting our prompts, we can mimic the styles of well-known artists through the use of specific language. We can use the describe command, render the results, and do some tweaking to modify and manipulate using the best set of terms Midjourney provides to us.

However, it's equally apparent how this method can lead to wildly inaccurate results—sometimes leading to delightful surprises, other times, not so delightful deviations.

On trend with our previous two tests, let's see what we get if we simplify our prompt and just enter in the Author and work title for reference. I simply entered in "/imagine Hilma af Klint's The Swan, No. 17" to see if Midjourney would replicate the art in question better or worse than when it was given the image to describe first.

In this test, different from our previous two tests, we get a wonderfully inaccurate set of images when we do this! I can't help but laugh! Which one is your favorite? I quite like the first one a lot!

More Experiments with AI Image Generators & Art History on the Way!

Stay tuned for Part 4, where we'll explore the work of another famous artist and see if Midjourney can replicate it! If you enjoyed this experiment, you can also check out the previous two experiments. Check out the Midjourney AI Comparisons of Sandro Botticelli or the Midjourney AI Comparisons of Paul Cezanne.