Every once in a while a new geospatial tool shows up and you can tell immediately it has range.

That was my reaction to forge3d, a new Python package by Milos Popovic: a GPU-accelerated 3D terrain renderer built in Rust and shipped as prebuilt Python wheels. No Rust toolchain and no elaborate setup: you type pip install forge3d and go.

That’s already a pretty good pitch. What got me interested, though, was the promise of a terrain workflow that felt fast and modern and at home in Python. So I did what I usually do when I find a new library that looks promising: I spun up an isolated pixi environment, opened a few notebooks, and started poking at it until I understood the shape of the thing.

I was up and running much faster than I expected.

Mount Rainier, Washington β€” terrain render showing 4,391m summit with terrain colour and hillshade, rendered with forge3d and matplotlib

Mount Rainier, Washington Β· 4,391m summit Β· 30m DEM Β· terrain colour + hillshade Β· rendered with forge3d + matplotlib

Why I Tried It

Washington State has very good open LiDAR data, so it was the obvious place to start. I mostly wanted to see what forge3d could do with lighting, relief and terrain shading. One of the nicest surprises was how quickly I got to something usable.

I had been doing similar terrain work in R with Rayshader, and this was much faster to get off the ground. Once I understood the basics, it opened up fast. After that the software wasn’t the bottleneck anymore. My own imagination was, plus however much good open LiDAR data I could find.

That’s probably the best compliment I can give a new geospatial tool: it got out of the way fast enough that I could start thinking about images instead of setup.

First Impressions

What stood out right away was that forge3d seems aimed at people who make maps, and not only at graphics people. Even this early, it already has the bones of a serious terrain workflow: DEM loading and terrain rendering, a large colormap catalog, CRS and reprojection support, Jupyter support, vector overlays, labels, and point cloud handling, and a Python API that is readable enough to learn from directly.

That last part matters more than people admit. Plenty of geospatial libraries technically work but make you fight them. This one seems to want you to get from raw elevation data to a render worth keeping with as little ceremony as possible, which is a good instinct for a mapping library.

I Started With Mount St. Helens

I didn’t start by trying to reproduce the polished hero renders from the project page. I started with Mount St. Helens.

Partly because it’s one of the most recognizable volcanic landscapes in the Pacific Northwest, and partly because it’s a great terrain subject. Between the crater, the blast zone, the radial drainage and the contrast of broad landforms with fine surface texture, it gives you a lot to work with right away.

It was a good test case for another reason, too: Washington's LiDAR data is good enough that you can see the difference as you push the resolution. That led me to make three images: a 10-meter vs 1-meter resolution comparison, a standalone annotated terrain plate, and eventually a render of Crater Lake just to see how the workflow handled a completely different landform.

Mount St. Helens, Washington β€” terrain render showing the crater and blast zone, USGS 3DEP data, rendered with forge3d and matplotlib

Mount St. Helens, Washington Β· summit 2,526m Β· USGS 3DEP 10m Β· rendered with forge3d + matplotlib Β· The crater and radial drainage make it a great first test subject.

The Resolution Comparison

Same mountain, same color scale, same basic framing, and still the jump from 10-meter USGS 3DEP data to 1-meter Washington DNR LiDAR changes the image completely. Gullies sharpen. Flow structure appears. The surface texture reads much more directly.

That’s where the package stops being "cool rendering tech" and starts being useful. The difference isn’t subtle: one render shows you the landform, and the other starts to show you the geology.

Mount St. Helens resolution comparison β€” USGS 3DEP 10m on left versus WA DNR LiDAR 1m on right, same area, same colour scale

Same area Β· same colour scale Β· 10Γ— difference in terrain detail Β· USGS 3DEP 10m (left) vs WA DNR LiDAR 1m (right) Β· Washington's open LiDAR data is where this workflow started to feel especially interesting.

Then Crater Lake

After St. Helens I wanted to test something structurally different, a caldera instead of a cone. Crater Lake works for the same reasons St. Helens does: a dramatic landform with a clear story, and terrain that earns its elevation data.

The USGS 3DEP data handled the caldera rim and the deep lake basin cleanly, and Wizard Island stands out in the render. At this elevation range the color scale does what you want, separating the water, the rim and the surrounding plateau without any hand-tuning.

Crater Lake National Park, Oregon β€” terrain render showing the caldera, Wizard Island, and surrounding volcanic landscape, USGS 3DEP data

Crater Lake National Park, Oregon Β· deepest lake in the US Β· rim elevation 2,716m Β· USGS 3DEP Β· rendered with forge3d + matplotlib Β· Wizard Island reads clearly in the basin.

My Favorite Output Wasn't About the GPU

Funny enough, the render I liked most was built with a pretty plain stack: forge3d for the terrain workflow, NumPy for the hillshade-style processing, and matplotlib for composition and output. There was no interactive viewer, no camera flythrough and no complicated rendering pipeline. Python, elevation, light and composition, and that was it.

I count that as a point in forge3d’s favor. Good tools don’t force you into one look or one workflow. forge3d seems built for high-end interactive and rendered output, but it also works as a clean way into terrain data and scene-making, even when the final image goes through the Python tools you already trust.

What Stood Out

Packaging that makes experimentation easy. A Rust-backed rendering library shipped as Python wheels is easier to say yes to. There’s a big difference between "install this and start playing" and "clear your afternoon because your build chain might break." Making it easy to try things is exactly what a young library should do.

It feels designed for GIS people. It doesn’t read like a graphics engine pretending to understand maps. The examples, the terminology and the data model all suggest somebody who knows what geospatial users want to do: load terrain, manage coordinate systems, bring in overlays, place labels and make strong images.

The source is readable. I like learning new libraries from the source while the docs are still catching up, and forge3d’s code is clear enough that this is productive instead of intimidating. That matters early in a project's life.

Why I Think It's Worth Watching

A lot of geospatial Python tools are either analytically strong and visually awkward, or visually strong and disconnected from everyday GIS workflows. forge3d looks like it’s trying to be both.

Terrain visualization is still more scattered than it should be. You can do it in desktop GIS, in game engines or in custom graphics stacks, and you can brute-force it in matplotlib. There’s still room for a tool that hands you a modern renderer and a Python API, with a workflow that feels natural to geospatial people. That’s a real niche.

A lot of exciting geospatial tooling is coming online in 2026, and forge3d feels like part of that wave.

What I Want to Do Next

Now that the basic workflow runs, I want to push it harder. Next on my list is finding the highest-grade LiDAR I can get my hands on for Mount Rainier and seeing what 1-meter resolution does to that summit terrain. The 30m DEM render above already looks good. I expect the WA DNR LiDAR to turn it into something else.

Once I got the hang of forge3d, I was mostly limited by my own ideas and by the open data I could find, which is a fun place to be. For a first pass, I can’t give a tool a much stronger endorsement than that.

I installed it because I was curious. A few notebooks later, I had images I wanted to keep.

Update, September 2026: I finally made the camera flythrough. Once Around Humphreys Peak orbits Arizona’s highest summit with USGS imagery draped over the terrain.


Go deeper

🎧 Listen & watch

3D mapping, plotting, and printing with rayshader — rstudio::conf 2019 (Posit), 12 November 2019, about 21 min video. Tyler Morgan-Wall shows off rayshader, the R cousin I was using before forge3d.

#14 – Mike Zoeller and Adam Mosbrucker — LIDAR Magazine Podcast, 28 October 2024. Two USGS volcano geologists on how lidar is used in the Cascades and Hawaii, from landing helicopters to tracking sediment.

📄 Read

Horn, B.K.P. (1981). Hill shading and the reflectance map. Proceedings of the IEEE 69(1), 14–47.

The 1980 Eruptions of Mount St. Helens, Washington — Lipman & Mullineaux (eds.), USGS Professional Paper 1250, 1981; the big book on how the crater got its shape.

Geologic Map of Mount Mazama and Crater Lake Caldera, Oregon — Bacon, Ramsey & Dutton, USGS Scientific Investigations Map 2832, 2008; Wizard Island and the rim, rock by rock.

Top GIS Voices 2024: Milos Popovic — Atlas; a friendly profile of the mapmaker behind forge3d.

📊 Data & agencies

Washington Lidar Portal — WA DNR’s free 1 m lidar, the star of the comparison. · USGS 3D Elevation Program (3DEP) — the 10 m and 30 m elevation. · The National Map Downloader — grab the tiles yourself.

USGS: Mount St. Helens — Cascades Volcano Observatory’s page for the first test subject. · Crater Lake National Park — NPS.

forge3d — the code, with docs and examples. · rayshader — Tyler Morgan-Wall’s R package. · pixi — the environment manager.