wesbos/3d-model-anything

The problem

Getting a real room into a browser as geometry usually means either buying a photogrammetry pipeline and cleaning meshes by hand, or eyeballing dimensions from photos. Neither gives you a scene you can edit as code. This repo records a room with an iPhone (LiDAR depth plus photos), then hands that evidence to an AI coding agent guided by a skill file so the agent measures, models and iteratively compares renders against the source photos.

What's different

The interesting part is not the viewer — it is the skill at .agents/skills/scan-to-threejs/SKILL.md. It encodes a reconstruction loop: read the capture export, check measurements, build the scene, render, diff against photos, adjust. That turns "make me a 3D room" from a one-shot prompt into a repeatable process with feedback, which is the main failure mode of agent-assisted modeling. The output is a hand-editable Three.js project rather than a black-box mesh from a scanning app. The README is explicit that this is not one-click scan-to-model, and the demo's Scan and Overlay modes are disabled because original scans aren't shipped.

Who it's for

Three.js developers and creative coders who want a room or interior as editable scene code, and who already use an agent with local skill support. It also suits people doing architectural/interior visualization prototypes. It does not suit anyone without a LiDAR iPhone or iPad for capture, anyone wanting a finished, installable product (there is no App Store release yet), or anyone who needs outdoor/large-scale scenes — the README limits the workflow to rooms and interiors.

Cost of adoption

You need Xcode to build the capture app, a LiDAR-capable device, and Node.js 22.12+ for the viewer (npm ci then npm run dev). Helper scripts need Python 3, Pillow for photo comparison, and optionally NumPy and Open3D for depth reconstruction; capture scripts call the agent-browser CLI. The code and skill are MIT, but the office images and textures carry separate asset terms — check them before publishing your own demo. The app is source-available and not yet distributed on the App Store, so treat it as a workflow you assemble yourself.

Verdict

The repo is three days old with roughly 1.6k stars and about +522 per day, but it has not appeared on the daily trending board and has no Hacker News discussion — that pattern reads as a launch spike from the author's audience rather than sustained adoption. There is real substance: a working viewer, a specified capture export format, and a skill that documents the iteration loop. But the end-to-end path is long and partly DIY, and the demo is curated. Worth cloning to study the agent-skill pattern even if you never scan a room; adopt it for actual work only if you already own the hardware and are comfortable maintaining the pipeline.