If you’re wondering whether Nano Banana Pro can credibly integrate a view of Roman-era Jerusalem into the rewilded landscape from the last post, the answer is yes. I appreciate how the above image even cleared some of the area around the walls, as you’d expect from history. The structures inside the city walls are mostly too large, however.
Here the rewilded landscape is misleading—during the time of Jesus (which the above image depicts), the area around Jerusalem was less forested than this image suggests. The area included agriculture, roads, pasturelands, and other changes introduced by humans.
Below is my attempt at using Nano Banana Pro to convey this human activity. It regraded the whole image slightly, and the roads aren’t exactly right. I also don’t think the Hinnom Valley south of the city would have this much agriculture. The terraced agriculture is a nice touch, though, since I spent so much time getting rid of terraces in the original image.
Here was my prompt:
Right now, this Roman-era city of Jerusalem feels pasted on, because it is. Integrate the feel of the city so that it integrates into the rest of the landscape.
Also add ancient roads and small-scale agriculture (think wheat barley, olives, and vineyards), reducing the forested area. Don’t have agriculture immediately outside the city walls. Especially include cultivated olive groves on the Mount of Olives across the gully to the east of the city.
Add a few small structures and villages in the area outside the walls (isolated farmhouses, etc.) that are appropriate for the time.
Make sure there’s a way to get into the city from the west (left) near where the walls make a “J” shape.
Keep the rest of the landscape as-is and don’t adjust the overall lighting or colors of the scene, just of the city.
Nano Banana Pro can rewild photos of archaeological sites with AI; it can also create rewilded maps. For example, here’s a fake satellite view of the Jerusalem area with all structures, roads, and anything human-created removed:
And georeferenced in Google Earth:
AI enables creating this kind of map in a few hours, rather than the weeks it would have taken using traditional methods.
The effective resolution of this image is about 1.2m per pixel, equivalent to a high-resolution (and therefore expensive) satellite photo. (A true satellite photo would show mostly urban development here, of course, and wouldn’t be terribly useful for visualizing the underlying landscape.) The topography is mostly accurate; the vegetation coverage is speculative.
Methodology
First, I needed a relatively high-resolution topography for the area around historical Jerusalem: approximately 2.3km by 2.3km (about 2 square miles). The highest-resolution free Digital Elevation Models are 30m per pixel, which at this latitude gives a grid of about 100 x 100 elevation pixels. While that may not sound like a lot, it’s enough to create a final 2,048 x 2,048-pixel image—but the low resolution of the source data also reinforces how much the AI is inventing fine surface detail.
I started with the GEDTM30 global 30m elevation dataset (which, as a DTM, aims to give bare earth elevations, excluding buildings and landcover). Using these instructions, I created 5m contour intervals in QGIS and exported them to a png. I compared these contours with 5m GovMap contours; they differed in some details but were plenty close enough for this purpose.
Here’s where Nano Banana Pro came in. I gave it the contours and the following prompt (the “text” in the prompt refers to the contour elevation labels):
This is a detailed map of the area around Jerusalem. Convert it to an overhead aerial view. Preserve all the topography exactly. Remove all text. Apply landcover (especially trees and scrub) in a naturalistic fashion and show bare dirt, light scrub, and trees where hydrologically appropriate.
Smooth out all the elevation lines—there are only smooth hills, no terraces or cliffs. Use the elevation lines as a reference, not to create terraces. No terraces should be visible at all; just smooth them out.
The idea is to make it look natural, without any human developments.
As you can tell from my pleas in the prompt, Nano Banana Pro really liked making terraces (since the contour intervals look like terraces). I ended up generating twenty-four iterations but used the seventh one because it preserved the topography of the City of David especially well. Each generation had different pluses and minuses—some were better at color, some at vegetation, and some at hydrology. That’s part of the beauty of using AI: it allows rapid iteration and many generations at low cost. This project cost about $5 in total.
I also explored giving it a version of the DTM itself (with the elevations scaled to grayscale values 25 through 244), as well as a hillshaded version. Nano Banana Pro gave me roughly comparable results for each, but I preferred how the contour versions turned out.
With a 2,048 x 2,048-pixel png in hand, it was time for Photoshop. I used the spot healing brush extensively to remove visible terraces. I also went back to Nano Banana Pro to generate trees and scrub for certain areas, brought in parts of other discarded generations, and used Photoshop’s built-in generative features in some places. You can definitely see artifacts from my editing if you look closely at the finished map. I also added an exposed rock (just visible under the “m” in “Temple” in the above map) where the Dome of the Rock now stands.
Then it was off to Illustrator to add the text and the outline of the city walls. ChatGPT gave me a few pointers to refine the look.
Finally, I georeferenced the map in Google Earth and consequently adjusted some of the wall placement in Illustrator to align the wall more precisely with structures that are still visible today.
Discussion
I’ve never used an AI + real data workflow like this one before. It would’ve been prohibitively time-consuming to create this map without AI, which is part of the ethical question around using AI. Did I “steal” the hundreds or thousands of dollars I might otherwise have paid a cartographer-artist to create this map? More realistically, I never would have created it at all.
The map’s high degree of realism could lead people to believe that it reflects reality more than it does; at first glance, you could easily take it for a real satellite photo. The landscape that it depicts never looked exactly like it does in the map. This combination of extreme realism with plausible hallucinations captures the current state of AI in a nutshell: it looks real, but it isn’t.
The map depicts a pre-human landscape (thus the “rewilding”). Biblically, it’s closest to how it might have looked in Abraham’s time, before subsequent urbanization. But even during his time, there still would be settlements, visible footpaths, grazing areas, small-scale agriculture, and potentially less forest.
Nano Banana Pro’s interpretation of the elevation data is reasonable. I feel like it made some of the eastern hills ridgier than they are in reality, however.
It also did a good job with the trees and scrub, though they’re much more speculative than the topography. I chose, artistically, to forest the western half of the map more than the eastern half, since Jerusalem approximately marks where denser vegetation in the west would yield to sparser vegetation in the east. I may have gone too far in either direction—too much forest in the west and too little vegetation in the east.
In addition to reconstructing archaeological sites from photos, Nano Banana Pro can do the opposite: it can rewild them—removing modern features to give a sense of what the natural place might have looked like in ancient times. Where reconstruction involves plausible additions to existing photos, rewilding involves plausible subtractions from them. In both cases, the AI is producing “plausible” output, not a historical reality.
Mount of Olives
For example, the modern Mount of Olives has many human-created developments on it (roads, structures, walls, etc.). My first reaction to seeing it in person was that there were a lot fewer olive trees than I was expecting, and I wondered what it would’ve looked like 2,000 years ago.
Nano Banana Pro can edit images of the Mount of Olives to show how Jesus might have seen it, giving viewers an “artificially authentic” experience. It’s “authentic” by providing a view that removes accreted history, getting closer to how the scene may have appeared thousands of years ago. It’s “artificial” because these AI images depict a reality that never existed, combined with a level of realism that far outshines traditional illustrations. Without proper context, rewilded AI images could potentially mislead viewers into thinking that they’re “objective” photographs rather than subjective interpretations.
Rewilded Mount of Olives
The first image below is derived from a monochrome 1800s drawing of the Mount of Olives, which allowed Nano Banana Pro to add an intensely modern color grading (as though post-processed with a modern phone). The second is derived from a recent photo taken from a different vantage point.
Similarly, here’s Mount Gerizim, minus the modern city of Nablus. Nano Banana Pro didn’t completely remove everything modern, but it got close. If I were turning it into a finished piece, I’d edit the remaining modern features using Photoshop’s AI tools (at least until Google allows Nano Banana Pro to edit partial images).
This process only works if existing illustrations or photos accurately depict a location. If I owned rights to a library of photos of Bible places, I’d explore how AI could enhance some of them (with appropriate labeling), either through reconstruction or rewilding. A before/after slider interface could help viewers understand the difference between the original photos and the AI derivatives, letting them choose the view they want.
Restoration (using original or equivalent materials to restore portions of the original site) is another archaeological approach that AI could contribute to, but the methods there would be radically different.
Nano Banana Pro did its best job at converting the Mount of Olives illustration, in my opinion. I wonder if doing multiple conversions (going from a photo to an illustration and then back to a photo) could yield consistently strong results.
Nano Banana Pro does a plausible job of turning a real photo of an archaeological site into what the photo might have looked like if you’d taken it from the same vantage point thousands of years ago. You can imagine an app running on your future phone that lets you turn your selfies at historical sites into realtime, full-blown reconstructions (complete with changing your clothes to be historically appropriate).
Here’s a reconstructed view of Ephesus (adapted from this photo by Jordan Klein). I prompted it to add the harbor in the distance, which no longer exists in the modern photo.
Leviticus probably isn’t your favorite book of the Bible, with its long lists of cleanliness regulations and priestly procedures. But I’ve long thought that the natural format for Leviticus is the flowchart: do this, then this, then this. A flowchart makes the prose much easier to follow. So I spent about thirty minutes a week over the past year turning Leviticus into a series of flowcharts by hand.
However, with Nano Banana Pro, I was able to make more progress in an afternoon than I had in a year—going from raw Bible text to finished flowcharts in four hours. I didn’t even use any of the work I’d done over the past year.
Here are some examples of finished flowcharts:
Methodology
I first generated some test flowcharts to get a visual style I liked. I wasn’t planning on the illustrations being so friendly, but Nano Banana Pro came up with a clear and pleasing style, so I went with it.
My first thought was to display all the Bible text—NBP could actually handle it—but the summary view I ended up with was easier to follow, visually.
From there, it was mostly a matter of choosing logical verse breaks for each flowchart, which ChatGPT helped with. I then used this prompt and gave it a previously generated flowchart as a style reference:
Create an image of a flowchart for Leviticus [chapter number] (below). Use the image as a stylistic model. Match its styles (not content or exact layout), including text, arrow, box, and imagery styles. Structure your flowchart so that it fits the content. Integrate the images into the boxes themselves where appropriate; they’re not just for decoration. Present a summary, not all the text. Indicate relevant verse numbers, and include the specific verse numbers in the title, not just the chapter number. Never depict the Lord as a person.
[Relevant Bible text]
Often it took two or more tries to get the look I wanted, or to ensure that it got all the logic right. I originally wanted to have all the clean/unclean animals on one flowchart, for example, but I couldn’t get the level of detail I was going for. So they’re broken up by animal type into multiple flowcharts.
On the other hand, even when I forgot to adjust the chapter number in my prompt, NBP would still show the correct chapter number in the output—it knew the chapter I meant, not the chapter I said.
All the image resizing and metadata work on my side to prepare the final webpage was vibecoded. It wasn’t hard code, but it was even easier just to explain to ChatGPT what I wanted to do.
Discussion
These flowcharts are better than I could have executed on my own and only took about four hours to create, from start to finish. By contrast, my earlier, manual process involved taking notes in a physical notebook, and I’d only made it to Leviticus 21 after twenty hours of work. Turning those notes into a finished product would’ve taken perhaps another 100 hours. So I got a better product for 1/30 the time investment, at a cost of $24 to generate the images.
Those twenty hours I spent with Leviticus weren’t lost, as ultimately any time spent in the Bible isn’t. In generating these flowcharts, I already had an idea of what the content needed to be and that it worked well in flowchart form.
But still, I didn’t add much value to this process. Anyone with a spare $24 could’ve done what I did. I expect that people will create custom infographics for their personal Bible studies in the future—why wouldn’t they?
The main risk here involves hallucinations. NBP sometimes misinterpreted the text, and the arrows it drew didn’t always make sense. I reviewed all the generated images to cut down on errors, but some could’ve slipped through.
As you can tell from my recentblogposts, I think that Nano Banana Pro represents a step change in AI image-generation capability. It unlocks whole new classes of endeavors that would’ve been too costly to consider in the past.
In April, I had GPT-4o create a bunch of maps of the Holy Land based on an existing public-domain map. My chief complaint at the time was that GPT-4o “falls apart on the details”—it gives the right macro features but hallucinates micro features (such as omitting specific hills and valleys and creating nonexistent rivers).
Nano Banana Pro changes that. It preserves features both big and small and doesn’t alter the location of features you give it, which means that you can hand it a map, have it transform the look, and then export it back out of Nano Banana with the correct georeferencing. You can completely change the appearance of a map and just swap it out for your purposes.
This time, I started with the same public-domain map but had Nano Banana Pro extend it so that it would have the same 2:3 aspect ratio as the GPT-4o images. It did a phenomenal job. If you’ve heard of the “jagged frontier” of AI, this work is an example of “sometimes it’s amazing.” There’s no reason why it should be so good at creating a map this accurate. But here we are. (You can download the 4K version of the generated image.)
Then I ran the same prompts on Nano Banana Pro that I used for the earlier GPT-4o images. The results preserve all the details but apply the appropriate style. While the Nano Banana Pro images are more accurate, I feel like the GPT-4o images were, on the whole, more aesthetically pleasing for the same prompt. On the other hand, the NBP images followed the prompts way better. Only a few of the more heavily stylized NBP images inserted the nonexistent river between the Red Sea and the Dead Sea.
Compare the “shattered crystal” look between GPT-4o and Nano Banana Pro. GPT-4o is more conceptual, while Nano Banana Pro is more literal.Compare the “painter’s impression” look between GPT-4o and Nano Banana Pro. To my eye, the GPT-4o one captures Impressionism better.
Below are some of my favorite Nano Banana Pro images. The first two recreate the Shaded Blender look that’s so hot right now. The second two show how NBP can change up the style while preserving details. I especially love how the last one makes the Mediterranean Sea feel vaguely threatening, which captures ancient Israelites’ feelings toward it.
This image (made with Nano Banana Pro), recreates one of my favorite views of the Holy Land. The original (by Hugo Herrmann) dates from 1931 and is in the public domain. The use of forced perspective makes the topography of the region clear, especially the relationship of the Jordan rift valley to both the Mediterranean Sea (to the west) and the hilly terrain (to the immediate east and west). Mount Hermon in the far north makes clever use of the horizon line to show its dominance.
A view like this also illustrates why biblical writers talked about going “up” to Jerusalem (which is on the peak nearly due west from the northern end of the Dead Sea near the bottom).
The original uses an older style that’s less immediately accessible to the modern eye. Nano Banana Pro is the first AI image generator to do a good job at updating the original’s appearance while removing text and other modern features. Nano Banana Pro also preserves topographic details (which are stylized in the original and not completely accurate) amazingly well. You can tell that it’s AI-generated if you zoom in on the high-resolution version linked above, though—its details feel imprecise compared to what a human would create.
I wanted to have Nano Banana Pro draw Saul’s path from Jerusalem to Damascus using a map reference, but all its attempts were wrong in various ways. So it does have limits. But those limits probably won’t exist in six months.
Google this week launched Nano Banana Pro, their latest text-to-image model. It far outshines other image generators when it comes to historical recreations. For example, here’s a reconstruction of ancient Jerusalem, circa AD 70:
I gave it this photo of the Holyland Model in Jerusalem and told it to situate in its historical, geographical context. Some of the topography isn’t quite right, but it’s pulling much of that incorrect topography from the original model. It can also make a lovely sketched version.
It also does Beersheba. Here I gave it a city plan and asked it to create a drone view. The result is very close to the plan; my favorite part is the gate structure and well.
It was somewhat less-successful with Capernaum (below). I gave it a city plan and this photo of the existing ruins. It’s kind of close, though it doesn’t exactly match the plan. It’s almost a form of archaeological impressionism, where the image gives off the right vibes but isn’t precisely accurate. Also try a 3D reconstruction of this image using Marble from World Labs.
Finally, I had it create assets that it could reuse for other cities for a consistent look:
I then had it create a couple typical hilltop shepherding settlements using the assets it created (again using “drone view” in the prompt):
Last month’s release of GPT-4o’s image-generation capabilities led to a huge improvement in instruction-following capabilities—specifically, it can now make maps that (more or less) match real geography.
The results match what James Farrell found in his similar cartographic explorations: GPT-4o creates “generally accurate topography” but falls apart on the details. In these maps, for example, it really likes to connect the Dead Sea and the Red Sea with a nonexistent river. And it includes the Sea of Galilee only when it feels like it. The details of the topography itself—hills, valleys—are broadly correct but wrong in details.
It tends to do better at geographically accurate reproduction when it’s generating something close to what it likely saw in its training data. Sometimes modern features, like country borders, leak through into the generations.
This kind of “vibe cartography” is different from what JJ Santos describes when using a similar term, where you can use Claude to automate map creation inside QGIS. In that process, you should end up with geographically “correct” results, but you’d have to spend a lot of time to achieve the artistic effects in the more conceptual maps here.
Evan Applegate at the Very Expensive Maps podcast likes to say that “you should make your own maps.” I don’t know that he’d consider this process to be “making” a map so much as vibing it into existence. I can imagine a cartographer using an AI to explore a certain look and then polish and execute that look using a more-traditional cartographic workflow.
Methodology
I started by uploading to Sora the finest map of the Holy Land ever created, which is in the public domain, and using that image as a base. From there, I started with this prompt:
Turn this hand drawing of the natural vegetation and topography of the Middle East into something different while maintaining the physical features (especially note that everything south of the Dead Sea is desert; there’s no river), without labels, human features, or political borders:
And followed it up with the specific style, with wording suggested by ChatGPT. For example:
A pure, traditional Swiss-style shaded relief map of ancient Israel — delicate shading for terrain, clean coastline, classic colors, masterful light sourcing.
You can find all the prompts by hovering over (or long-pressing) the images on the AI Maps page.
Posted in AI, Art, Geo | Comments Off on Doing Bible “Vibe Cartography” with GPT-4o
They’re all basically the same concept, with a happy sheep coming toward the camera. Prompting for a video is different from prompting for an image; I struggled to get good results in the limited number of generations available to me. I had more failures than successes.
Here are a couple of fails where I tried to get a video of Moses parting the Red Sea. The first one looks like a video game cutscene, but revealing a giant wall is opposite of what I’m going for. In the second one, Moses decides to take a quick dip in the Red Sea before popping back out. Both of them are trying (and failing) to create the “wall of water” effect popularized by the movie The Ten Commandments.
If I had more credits available, I’d share more. We’re in the earliest days of text-to-video generations—the DALLE-2 era of AI videos: they’re amazing but limited, advanced but (in retrospect) basic.
Posted in AI, Art, Video | Comments Off on Making Short Bible-Story Movies with Sora
.jpg){kind=link}
.jpg){kind=link}
