reverse-cad

Recover a building's camera and parametric model from ordinary photos, and prove the model against measured drawings.

Try it in the browser: studio · demo

Videos: 30 s teaser · studio walkthrough, 2 min · results walkthrough, 65 s

Every number in a model says where it came from: P read off a drawing or a published dimension, D measured from a photo, I inferred by a design rule, A assumed. A model is done when it lands on the photo from the recovered camera and its D and I values agree with sources the solve never saw.

What is here

• Agent skills (skills/): reverse-cad runs the whole job, reverse-photo solves photos, reverse-plan reads measured drawings. Plain SKILL.md files for Claude Code or any agent that loads skills.
• A camera and parameter solver (tools/solve.py, tools/solve_multi.py): marks on the photo tied to 3D expressions over named parameters; camera and parameters solved together, from one photo or several at once, with bootstrap spread, hold-out and a sweep over camera assumptions.
• Four cases (cases/) checked against Historic American Buildings Survey drawings, with the drawings and photos in sources/ and photos/.
• A studio (studio/index.html): each solved building as a 3D parametric model (see below).
• A demo page (demo/index.html): move a parameter and watch the model leave the photo; hover one to see which marks measure it.
• Rhino output: a generator per case, the .3dm files, and a viewport capture aligned to the photo pixel for pixel.

Studio

Open it online, or open studio/index.html from a clone (three.js loads from a CDN).

Model: every element is a formula over named parameters; sliders are graded P/D/I/A, design rules such as the number of bays can be changed	Plan: dimensions in metres, recovered from the photo; drag a grip and the plan, the 3D model and the checks follow
Graph: which parameter moves which element, found by perturbing each one	Photo: the camera flies back to where the photographer stood and the model lands on the photo; the panel shows the error in pixels and every check against the drawings
Review: the model's history and proposed changes from colleagues or agents, a 3D diff (moved amber, added green, removed red) and design conflicts in plain words, with Accept and Reject. No git vocabulary needed	Robie House: one photo fits itself and misses the drawings; two photos solved together land on them

The review screen is the interface side of paramgraph, which does the same diff and merge in the terminal and in CI.

Results

From BENCH.md, generated by tools/bench.py:

building	photos	within 5 % of the drawings	worst
Farnsworth House (Mies, 1951)	two, solved together	9 of 9	4.1 %
Robie House (Wright, 1910)	one iPhone photo	0 of 3 walls	18 %
Robie House	two, solved together	6 of 7	5.4 %
Gamble House (Greene & Greene, 1908)	two, solved together	3 of 10	14.5 %
S. R. Crown Hall (Mies, 1956)	one	column bay −1.0 %, ceiling +0.1 % against published figures

The Robie House rows are the main lesson. One photo fitted itself to 2.5 px, its hold-out and camera sweep were quiet, and it was 11–18 % off the drawings. Only a second photo caught it; solved together, those three values came within 0.1–1 %. Gamble House shows the current limit: with layered trim and pitched roofs, deciding which edge a mark refers to matters more than the solver.

Quick start

Python 3.10+ with numpy, scipy and opencv-python. Run from the repository root.

# open in a browser (the studio loads three.js from a CDN)
studio/index.html
demo/index.html

# solve one photo, with bootstrap spread and hold-out
python -I tools/solve.py cases/farnsworth/spec_9970.json --boot 100 --holdout

# solve two photos of one building together
python -I tools/solve_multi.py cases/robie/spec_w1.json cases/robie/spec_w2.json --boot 30

# what each parameter depends on, before solving
python -I tools/dag.py cases/robie/spec_w1.json

# the benchmark table and the synthetic tests
python -I tools/bench.py
python -I -m pytest tests -q

A spec is JSON: parameters (free or fixed, with notes), the camera's starting guess, point and line marks with their 3D expressions, and reference values kept apart as truth. See the docstring of tools/solve.py and the specs in cases/.

To use the skills, copy skills/* into your agent's skills folder (for Claude Code, .claude/skills/ in a project or ~/.claude/skills/).

Models as text

tools/to_pg.py writes a solve as a plain-text parameter graph for paramgraph: parameters with their grades, rules, boxes and the drawing values as checks. In git, a change to the model then reads as a design diff, and merges stop when two branches change things one element depends on. docs/git-for-bim.md is the design note for reviewing changes to shared building models this way.

Limits

• Marks are read by hand on zoomed crops (tools/grid.py), with edge detection to number and refine them (tools/segments.py, tools/snap.py). This is the slowest and most error-prone step.
• Geometry is boxes, lines and gables; no curves.
• One photo gives proportions; scale comes from one known dimension per case.
• Reference values for Gamble House are from the 1908 design drawings, not a survey of the built house.

Credits and licence

Code and documentation: MIT, see LICENSE. Photos keep their own licences, listed with their authors in photos/CREDITS.md. The HABS drawings are public domain; sources for each building are in sources/<building>/SOURCES.md.

Part of vibecading, an open lab for AI-native tools in architecture, 3D and games.