Generative Archaeology: AI-created realistic images of ancient worlds

Motivation

History is fascinating, but it’s difficult to visualize what the world looked like thousands of years ago.

Many aspects contributed to this difficulty. Partially it’s because ancient people had a different sense of aesthetics. Before the 1400s, painters rarely had perspective in mind, making proportions of things look rather odd to modern eyes. Partially it’s because time had rotten away a lot of details from the artifacts. Potteries, sculptures, and buildings have faded in color over time. For example, did you know that Egypt’s Great Pyramid of Giza was white and had a gold cap? All these factors can make it hard for people today to appreciate the beauty of ancient times with accuracy.

Recently, I’ve learned about people using AI to create images for scenes in the past. I call this trend “generative archaeology”.

• On the scale of decades, some street photographers have been using AI to create photorealistic images of China in the 1990s.
• On the scale of centuries, some artists have been creating images of important historic scenes (such as the Russian Revolution) and important people (such as Joan of Arc) with a dramatic touch in the styles.

I thought to myself: Why not extend that further and go several millennia back in time?

A map of ancient Mesopotamia

I’ve always been interested in Mesopotamian history, so I decided to start there. Several civilizations roamed that region: Sumerians, Assyrians, Babylonians, Akkadians, and so on. Where should I start?

While exploring places as a tourist, I’ve always found a pictorial map helpful. In simple words, a pictorial map ditches cartographical accuracy; in exchange, it illustrates important points-of-interests, such as tourist attractions, local plants and animals, and important landmarks. Let’s see what MidJourney can offer us. Here is my prompt:

A pictorial map of ancient Mesopotamia around 2000 BC, slightly tilted like an aerial view, featuring important cities such as Ur, Uruk, and Babylon. The rivers Tigris and Euphrates are clearly visible. Rivers and cities are labeled in English and in cuneiform text.

After a couple of refinements, this is what I arrived at:

Of course, you shouldn’t trust this map for your next time-traveling adventure. Flaws are quite apparent:

• There’s no legible label, either in English or in cuneiform.
• Nothing says Mesopotamia in this map. I was expecting at least one ziggurat per town.
• Architectures and sails look like alien tech. What even is this?

So much of the aerial view of Mesopotamia. What about a view in the city?

A typical market in Ur

Ancient Mesopotamian art were mostly for the gods or for the royalties, especially sculptures. I wonder what a typical day of ordinary people would look like.

Here’s a scene I generated with MidJourney:

Photo of a busy marketplace in the ancient Sumerian city of Ur around 2039 BC. In the foreground, a diverse group of people are engaged in various occupations, including a scribe writing cuneiform on a clay tablet, 3 soldiers patrolling the area, a farmer walking by, and several merchants selling handicrafts. Around the street corner, there is a pub. Beside the window, someone is sipping beer from a straw. In the distance, you can see the ziggurat. People wear traditional ancient Mesopotamian garments appropriate for their occupations.

What a buzzing town! You can almost hear the noise from the animals and the people in this picture. Kudos to MidJourney, it did a great job including these details:

• The AI included a goat in the foreground. Mesopotamians have already domesticated goats for 6,000 years by the time illustrated here.
• The street is paved. This is also fairly accurate — Ancient Mesopotamians are believed to be the first civilization to have paved roads.
• Windows are grilled with wooden strips. Since wooden stuff usually has rotten away in archeological sites by the time they are uncovered, it’s fairly easy to forget about them. The fact that MidJourney remembered to include them is worthy of praises.

There are, of course, several areas that could be improved:

• Even though I requested a photo, the AI generated a painting.
• Where’s the beer buddy that I specifically asked for? Mesopotamian beer was so cloudy with breadcrumbs that people had to sip the drink with reed straws, as depicted in this video from the Mists of Time team. (Hint: You can brew ancient Sumerian beer from home.)
• What’s that guillotine-looking thing in the distance, on the top of a building? Looks like a safety hazard to me.

Besides the clear wins and losses, there are also a couple of points that I am uncertain about:

• Buildings are mostly two levels high and flat on top. This is quite accurate, as attested from this terracotta model of a house from 2,600 BCE Babylon. However, I’m suspicious about the under-maintained state of the houses. Shouldn’t they be in better condition when still occupied?
• Clothing. It’s true that tunics and turbans are common in ancient Ur, but I was expecting more fringes, honestly.

Much of the joy of generating stuff with AI lies in the randomness. You can have an infinite (although limited by your budget) number of illustrations on the same topic. Here are a couple of other scenes I generated with the same prompt:

In this illustration, we see more types of livestocks, including sheep (bottom-left) and cattle (bottom-right).

This edition features several different details:

• Pots and baskets on the top of the ceiling (middle-left) and on the ground (bottom-right). They seem to be holding mushrooms of sorts.
• Strands of grains hanging from various places. Looks like a year of harvest!
• Windows are smaller than those in the previous two paintings, which is perhaps more accurate.

I’ll leave the rest of the paintings to you to explore. See if you can confirm or debunk more details of the paintings.

Aesthetically pleasing though these paintings are, you may argue that they merely depicted a generic scene common to all middle eastern marketplaces. Can we emphasize the ancient Mesopotamian element?

Interior design, but it’s ancient Assyria

In 1854, Assyriologist Austen Henry Layard drew an illustration for a hall in the Assyrian Palace of Ashurnasrirpal II:

This is perhaps the most detailed (and credible) painting of the interior of an Assyrian building we have to date. I fed it to MidJourney and let its imagination fly:

Repaint this room in DSLR photography. This is a hall in the Assyrian Palace of Ashurnasrirpal II. This is from the ancient Assyrian Empire. Similar to Babylonian, Akkadian, Sumerian cultures, they are all from the Mesopotamia.

After hand-picking the images for ~5 iterations, I arrived at this picture:

So many pillars/columns! The columns shown in this scene seem to me like a good mixture of Egyptian- and Persian-style, with a generous application of blue color. This color reminds me of the glaze of the Ishtar Gate, which should have looked quite shiny under sunlight. In antiquity, the gemstone lapis lazuli is well celebrated for its shiny blue color. Considering the scarcity of lapis, ancient Assyrians may have associated the color blue with high social status, thus explaining the usage of blue in palaces. In comparison, Layard’s painting didn’t include any pillars, so this is a creative invention of the AI’s.

Just like Layard’s drawing, this generated artwork depicted reliefs on the walls. Reliefs are a common form of sculptural art in ancient Mesopotamia. Specifically, I like how the AI put — above the relieves on the left-side wall — a second row of relieves (on the wall facing towards us). I wish the figures could’ve been more legible, though.

Bringing a relief to life

Speaking of relieves, now that we are at it, why don’t we try to generate a realistic depiction of someone from a relief?

Safeguarded in the Met, this relief from the Ninurta Temple depicts a man with wings, perhaps also a deity, whose name is unknown:

I asked MidJourney to visualize him as a real person:

This person but in a real-world photo. He has wings. He holds in his left hand part of a plant whose three branches end in rosettes. This figure is from the ancient Assyrian Empire. Similar to Babylonian, Akkadian, Sumerian cultures, they are all from the Mesopotamia.

This is quite good! The facial features have been captured quite nicely. He also looked more authoritative:

• Instead of holding a wilted rosettes in a down-pointing left hand, he is now raising the plant up high like a scepter.
• His clothing is more symmetric (though the AI has once again failed to capture the frills ).

The whole thing still looked like a sculpture, though. If you look at his wings and the outfit, it’s hard to imagine how he could fly or even walk. Nevertheless, this is a huge leap forward from what he used to look like in the Temple.

Afterword

A brief recap. In this post, I presented some AI-generated artwork about ancient Assyria. From the example of the aerial map, we saw that the AI failed miserably at tasks that 1) require attention to too many details and 2) lack adequate training data (I suppose pictorial maps weren’t very popular during ancient Mesopotamia). On the other hand, it does an excellent job in themes that are well-explored by artists throughout history (that is, the marketplace), scenes with fewer details (i.e., interior), and portraits of single persons.

Plan ahead

Next, I plan to expand on the idea of generating portraits of ancient people. For example: In the 1882 lookbook Costumes of All Nations, on page 4 in the chapter “Ancient Times”, a scene of pre-battle Assyrian military was depicted:

This illustration is quite realistic, but wouldn’t it be cool to see real people wearing these costumes?

There’s also the Feminine Costume of The World by Paul Louis de Giafferri from 1886. In Part 5, Plate 2, several outfits were illustrated.

If you are interested in this topic, you should read Traditional clothing of Mesopotamia. What did it look like?, which features some excellent drawings of Sumerian and Assyrian clothes by Tadarida.

Significance

Great educational value lies in “generative archaeology”. Many people are interested in learning about ancient cultures, but the scarcity of illustrations and access to historical relics (think: excavation sites, museums, antique shops…) may hinder their curiosity. This is more serious for young people, for imagination requires experiences as ingredients!

With generative AI, things can change for the better. We can turn sketches to life-like pictures. For the ages where arts are reserved for the powerful, we can illustrate what it looks like to live as an ordinary citizen. Exciting possibilities lie ahead of us by introducing AI technology to history educational programmes.