Building Winsted twice: first as a map, then as a town.

The spark

A tiny-town repository made the idea feel possible.

The reference was TinyTown: open terrain, streets, building footprints, and hand-touched landmarks assembled into an explorable miniature. The attractive part was not “make a map.” It was the possibility of recognizing a place because its ordinary buildings felt specific.

“Can you use this and develop a 3D map of Winsted, Connecticut?”

The first definition of done was deliberately small: make Winsted load, make it navigable, and make ABC Pizza and Dairy Queen recognizable enough to prove the local-landmark idea.

Open-data foundation

The town arrived as geometry before it arrived as character.

OpenStreetMap supplied roads and building footprints. Open elevation data shaped a roughly five-kilometre landscape around Winsted. The initial scene contained 3,553 buildings—enough to establish the town’s density and street pattern, but mostly as simple extrusions.

That first pass answered the large-scale question: Winsted could exist as one continuous Three.js scene. It also made the fidelity problem obvious. A correct footprint with a generic roof is geographically useful, but it is not yet the building somebody remembers.

Local anchors

ABC Pizza and Dairy Queen became the first test of “that’s our town.”

ABC Pizza and Dairy Queen were not selected because they are architecturally grand. They were selected because they are staples. Their silhouettes, roofs, storefront proportions, and signs carry more local recognition than another generic block.

The models were authored as lightweight web geometry so they could sit inside the complete town without turning two landmarks into a performance tax. That tradeoff—specific enough to recognize, simple enough to stream—became the project’s central design rule.

Load the live 3D town

The actual published map

Load it here to orbit, or open it full-screen. The preview stays still until requested to protect mobile memory.

Open Winsted in Miniature

The phone said no

The desktop scene worked. Mobile exposed the real delivery problem.

The first public build tried to load too much town at once. On a phone, the experience could stall or fall back into expensive local generation. The fix was architectural rather than cosmetic: bake a small base, divide the town into tiles, then request detail near the camera.

The production manifest now describes 2,231 tiles, 168 regions, and 225 detailed tile assets. Winsted defaults to this streamed path. The public experience looks like one town, but it arrives as a sequence of bounded decisions.

Finding a place

A model of a whole town still needs the most ordinary map feature: “take me there.”

The address layer came from Connecticut’s public address-point dataset. It is kept separate from the 3D building geometry, then projected into the same local coordinate system at runtime. Search accepts street fragments and familiar place names, groups repeated source points, and calculates a camera view that fits the destination on desktop or mobile.

The shipped index contains 3,465 public address points. The UI turns those into a smaller set of searchable destinations without placing labels over every roof. That distinction matters: the address data overlays the experience, but it does not have to clutter the scene.

The LiDAR turn

Then we found the state’s actual 3D building shells.

Connecticut’s 2023 LiDAR-derived building dataset changed the fidelity ceiling. These were not footprints extruded to a guessed height. They carried rooflines, connected masses, and measured vertical form. A bounded pilot extracted 34 source buildings around nine landmark areas.

The gain is visible immediately: roof steps, annexes, and awkward real-world massing appear without hand-modeling each face. The limitation is equally important. These are geometric shells, not finished game assets. They do not contain storefront photographs, clean materials, or a solved terrain datum.

Three-building trial

ABC, DQ, and The Gilbert School became the first asset pipeline.

Three landmarks were isolated as portable GLB files. ABC Pizza became the compact commercial test. Dairy Queen preserved two separately identified adjoining source pieces because their ownership relationship was still ambiguous. The Gilbert School tested a larger, more complex footprint; its selected state shell overlaps the current main OpenStreetMap school footprint by 93.58 percent.

The geometry is lightweight enough for the web and traceable back to official Connecticut source IDs. Centering each asset helps inspection; it does not rewrite the source geography.

Local AI texture lab

Meshy looked convenient. The local open-source route taught us more.

A Meshy upload package was prepared, but no paid job was submitted. Reference-photo licensing was a real constraint, and the free workflow did not offer the control this experiment needed. The alternative became Blender 5.2, the StableGen add-on, ComfyUI, RealVisXL, and a depth ControlNet—running locally on the Mac.

The first prompt failed in a useful way. It asked for a “material texture” and a burgundy awning. The model read the shallow building silhouette as two enormous awnings attached to a brick wall. The second pass reframed the subject as an entire detached low-rise building. The third gave the overhead camera its own roof-only instruction.

Where this goes next

The realistic city is a hybrid, not one magic model.

The best path combines systems instead of asking one of them to do everything. OpenStreetMap keeps complete coverage. LiDAR supplies measured massing and roof shape. Address points make the town searchable. Licensed street-level photographs provide observed colors, windows, doors, signs, and material breaks. Blender handles alignment, cleanup, UVs, façade masks, and deliberate exceptions. Generated texture tools can fill low-risk gaps, but every landmark needs a truth check.

• Align the vertical datum.Prove each shell meets the same terrain surface before replacing production geometry.
• Build a reusable façade-photo kit.Capture or license front, side, roof, and corner references with clear provenance.
• Separate roof and wall materials.Prevent roof cameras from painting windows and wall cameras from inventing roof detail.
• Integrate by confidence.Replace a production building only after geometry, texture, scale, mobile weight, and local recognition all pass review.

The public miniature is already useful. The research pipeline is about earning the next level of realism without losing performance, provenance, or the completeness that made the first map work.