I build small 3D scenes of industrial sites as a way to study layout. It is cheaper than flying somewhere, and it forces me to notice how infrastructure is arranged. Palo Verde is a good stress test: three reactors, a very large switchyard, a water system that wraps around the site, and almost no vertical clutter to hide behind. This is a build log, not a product review. No drone, no CAD files, no survey. Just public imagery, a map selection, and a lot of orbiting.
Scope and constraints
The goal was a walkable massing model, not an engineering asset. I wanted to answer questions like: how far is the switchyard from the reactors, how much of the site is water infrastructure, and where does the transmission corridor enter the property. Those questions do not need centimeter accuracy. They need correct relationships. I also set a rule: every feature in the model had to be checkable against public imagery. If I could not verify something, it stayed a question rather than a wall.
Picking the bounding box
GenAnywhere takes a rectangle on a map and returns a textured 3D model of the area inside it. The box matters more than any prompt. Too tight, and you lose the transmission corridor that explains why the switchyard is where it is. Too wide, and the plant becomes a small cluster in a sea of desert while the mesh resolution drops. I settled on a box that included the fence line, about a mile of desert to the west, and the first row of transmission towers.
First pass: what the generator infers
It inferred the big primitives well: domes, halls, rectangular cooling towers, roads, and parking areas. It inferred the switchyard as a grid of gray boxes, which is roughly right at a distance. What it cannot infer is function. A low building could be a warehouse, a pump house, or a security post. The model picks a plausible shape and moves on. That is fine for massing and useless for operations. It is also confidently wrong in places: a row of parking stalls became a long low shed, and a drainage channel became a road. Those mistakes are easy to spot from above, and a reminder that the output is an interpretation, not a copy.
Where it breaks
Three failure modes showed up. Thin structures such as pipe racks, fences, and light poles tend to disappear or merge into walls. Repetition gets simplified: rows of identical transformers or parking stalls collapse into a slab. Grade is flattened. The site is nearly level, but the water basins sit slightly below grade, and the generator tends to put everything on one plane. None of these are fatal for a layout study. All of them matter if you care about drainage.
Second pass: reference-driven correction
I exported screenshots of the first model, compared them with satellite tiles, and marked the deltas: a missing pipe bridge, a turbine hall that was too short, a switchyard pushed too close to the fence. Regenerating with those references improved the silhouette. The lesson is that this workflow is iterative. Treat the first output as a draft and budget time for two or three passes.
Time budget
The whole exercise took an afternoon: about forty minutes to define the box and run the first pass, an hour to compare it with reference imagery, and another hour for the second pass and screenshots. Most of that time went into looking, not generating. If I had tried to model every pipe and fence, I would still be at it. The constraint was useful. It kept the model at the scale of the questions I actually had.
What I would do differently
Start with a diagram. Before touching the map, sketch the site as five zones: reactor, turbine, cooling, water, and electrical. Then choose the box to include the transitions between zones, because that is where the layout logic lives. Finally, keep a margin of empty desert in frame. The space around the buildings explains the buildings.
Takeaway
A generated 3D model will not replace a site visit or a proper CAD file. It does something different: it makes spatial relationships cheap to inspect. For a site like Palo Verde, where the interesting information is distance and arrangement rather than detail, that is enough to learn something real. I came away with a better mental map, three useful screenshots, and a longer list of questions than I started with. That is a good outcome for an afternoon.
