Blog

Let’s say you want a high-resolution (1.2 meters per pixel) hillshade like this one of cliffs and hills to the west of the Dead Sea:

So that you can layer it over a satellite image (compare the original satellite image without hillshading added):

Or maybe over an idealized landscape with human features removed:

But all you have is a lower-resolution (30 meters per pixel) hillshade like this:

Nano Banana Pro can help you out, if you’re willing to accept that it’s making up all the details it’s adding to your lower-resolution hillshade and that your high-resolution hillshade looks nice but doesn’t necessarily reflect reality.

Here’s how I made the above hillshade and tiled it to cover about 3,000 square kilometers around Jerusalem.

Process

First, I used Eduard to create a 30m-per-pixel hillshade derived from the recent CC-BY-licensed GEDTM30. I gave the hillshade to Nano Banana Pro along with this prompt, repeating it a few times until I was satisfied with the result. I considered whether to go straight from the DEM to the final hillshade (which does actually work decently), but I wanted to take advantage of Eduard’s hillshading know-how. I also wasn’t confident that I could use the DEM for tiling.

Once I had an initial tile, it was mostly a matter of creating tiles that extended from existing tiles. I ran Nano Banana Pro repeatedly with this prompt, overlapping each tile by 248 pixels for a 2K tile and 496 pixels for a 4K tile (about 25 square kilometers) to ensure that the style and luminosity were consistent between tiles. Here’s an example tile overlap with high-resolution hillshade on the right and bottom sides of the tile.

I did experience some style drift, however; the hillshades got fainter over time.

This process worked great for hilly terrain; I almost never had to regenerate a tile.

For terrain with large flat areas, however, this process fell apart quickly. It often took several tries, plus adjusting the amount of overlap between tiles, to get a usable result. Typically, Nano Banana Pro wouldn’t match the luminosity of the surrounding tiles, or it would add distracting detail to the flat area. It was possible to get a decent result, but it required lots of human attention and tinkering—in other words, it wasn’t an automated process like the hilly terrain was.

If you look hard enough, you can find some tiling artifacts in flat areas (and a few in hilly areas). In practice, these tiling artifacts won’t be visible to map viewers since you’re likely draping the hillshade over some kind of background and reducing the opacity or increasing the gamma to keep the hillshade from overwhelming the viewer.

I didn’t use Photoshop on any of these tiles (though I did sometimes run a histogram match between the source tile and the result tile), but I probably would need to if I were to create more tiles for flat areas.

Results

In all, I created hillshades for about 3,000 square kilometers around Jerusalem, spending US$70 on Nano Banana Pro (2.3 cents per square kilometer, or 6 cents per square mile). That cost includes a lot of experimentation; at scale, with a mix of hilly and flat areas, the all-in cost is about 1.8 cents per square kilometer.

This area represents about 15% of the area of the full extent of ancient Israel (“Dan to Beersheba”), which means it would cost around $500 to create a full set of tiles. I stopped tiling when I exhausted my budget for this project (and my patience for regenerating flat areas).

Here’s the coverage area:

Discussion

As noted above, the resulting hillshade is plausible but fake—there’s no way any process can turn a 30m hillshade into a 1.2m hillshade and reflect reality.

Whether you want to use this method depends on your application. If you’re creating a fantasy map, you’re already two steps removed from reality, so this method can add some extra realism to your map. If you’re doing historical mapping, you’re one step removed from reality, as climate, landforms, and landcover have shifted over time.

This method shines where you’re pushing past the detail available in the lower-resolution hillshade and want to provide a crisper experience without presenting all the detail that’s available in the higher-resolution hillshade. The Good Samaritan images below show where I think this method works especially well.

The hillshade quality is pretty good. In general, the results are hydrologically consistent (rivers drain in the correct direction). It also captures the traditional hillshade look exceptionally well, in my opinion, and this process scales well in hilly terrain. The limiting factor in hilly terrain is cost, whereas the limiting factor in flat terrain is the time involved to revise tiles. In flat areas, it might make sense to retain the lower-resolution hillshade or to use a different super-resolution method.

In principle, it would be possible to create a model similar to Eduard’s U-Net approach that could go from low-resolution to high-resolution hillshades without involving Nano Banana Pro. I’m skeptical that it would handle drainage properly, but the bigger barrier is that Google’s terms of service preclude creating such a model.

Conclusion

To give you a practical application, here’s a closeup of the road from Jericho (where the two roads intersect on the right) to Jerusalem (which is off-map to the left). This road reflects the setting of the Good Samaritan story. Everything on the high-resolution map feels crisper and clearer thanks to imaginary AI detail.

First the lower-resolution map:

And then the higher-resolution map:

The source 30m hillshade and derived 1.2m hillshade are both available here for your use. You’ll probably want a GIS tool like QGIS to work with them; you won’t be able to just use them as-is in Google Earth.

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.

Data

You can download a jpeg of the map with and without labels. The unlabeled version is available as a geotiff for your own GIS applications. I also added both the labeled and unlabeled versions to the Map Overlays for Google Earth page, where you can download a KML to explore them in Google Earth.