The first model looked convincing and was wrong by about forty meters. The station sat under the right avenue at roughly the right depth, but when I laid a street map over it, the entrances landed on the wrong blocks. The mesh was fine. The reference frame was not.
That is the step map-to-3D work tends to skip. A generated scene can look correct alone and still be useless for measurement, sections, or overlays. Before adding an escalator, I had to decide where the model was, what zero meant, and which way was up.
What the source gives you
GenAnywhere turns a polygon I draw on a map into a massing model of the buildings and terrain. It is a scene, not a survey. There is no construction datum, no legal coordinate system, and no guarantee that the origin means anything. Those are my decisions.
The goal is not survey accuracy. The goal is a model that can answer simple questions without lying about scale.
Horizontal control
A web map is usually geographic, with coordinates in degrees. A 3D scene wants linear units. Converting between them changes the shape of the site slightly, and that difference matters when the model has to line up with a street grid.
For this site I used a projected coordinate system for New York State Plane, Long Island zone, and kept the scene in meters. I wrote down the projection and the local origin: the middle of the concourse, on the floor slab. The origin is arbitrary, but it is documented.
Vertical control
Vertical is harder. The street has one datum, the station another, and the drawings may use a third. The concourse is about 140 feet below Park Avenue, or 42.7 meters. That is a check, not a coordinate.
I set the concourse floor at zero and recorded the offset to the street. Then I checked the platform level against the known depth. When the massing came in too shallow, I adjusted the below-grade volume, not the whole scene. The street grid stayed fixed, because that is what a reader uses to orient the section.
Axis and bearing
Most 3D tools assume Y is up. Most GIS data assumes Z is up. A model imported without checking that convention will lie on its side. A model rotated to the wrong bearing will look almost right.
I tested the axis with a ventilation shaft that runs toward the street. If the shaft points sideways after import, the convention is wrong. I checked the bearing against the Manhattan grid and recorded the rotation. One degree is invisible in a screenshot and obvious in an overlay.
Three checks
Before adding detail, I verified three things: the doors fall between 44th and 48th Streets, the layout has eight tracks across four platforms, and the concourse sits about 140 feet down. If a model fails one, stop and fix the reference frame first. Detail makes a georeferencing error harder to see, not easier.
I also kept a short control-point list: a known street corner, the concourse center, and the top of the ventilation shaft. Each has a scene coordinate and a note. That list is worth more than a prettier mesh.
Keep the shell separate
The generated model is a shell: envelope and terrain, almost nothing underground. I put the platforms, escalators, tunnels, and equipment in separate collections and left the shell alone. If the shell is right and the fit-out is wrong, the error is local. If everything is merged, every fix becomes a rebuild.
What the tool cannot know
GenAnywhere does not know which door is open at night, which escalator is running, or how long a walk feels with a bag. It does not know construction sequence or the difference between passenger space and back-of-house. Those are survey, operations, and experience questions.
What it does give you is a credible massing in minutes. Georeferencing is what turns that massing into something you can measure and hand off.
The hour that saves the week
The fix took about an hour: projection, origin, axis, three checks, and notes. After that, every addition landed in the right place. The model stopped being a picture and became a reference. A section cut is only useful if the thing you cut is where you think it is.
