Casa Cornell
On this pageThe sequenceContents12 Where it stands
Preview build13 August 2026

A house rebuilt from two photographs of its floor plan.

The sequence below runs in this page. Drag the bar or pick a step. The room outlines, names and dimensions come from the same reviewed files the build reads. None of it is placeholder art.

Skip to the house, photographed → four views, every one held to the plan

Photograph of the first floor plan
The same plan after deterministic cleanup
Scanning
The house photographed from the south: cream
      clapboard, green metal hipped roof, full width porch on turned posts, gazebo at the left
Source photographCamera EXIF · 6 March 2026
00Contents

Five pages, in the order the work happened.

Each page holds one stage and stands on its own. The running index of every sealed run, payload directory, generator and gate behind them is in docs/INDEX.md.

plan.htmlThe plan

Paper, photographed at an angle under a window. What came off it, and what turned out to be wrong with what came off it.

  • Two phone photographs, and nothing else.
  • Classical computer vision, run to the point where it broke.
  • Fifty-one named rooms on the sheet. Three of them became geometry.
  • Every room in the reconstruction is the wrong size.
  • Eight degrees of freedom buy nothing.
house.htmlThe house

The sheets traced again from the numbers written on them, built into a solid that has a back, and photographed as the house it says it is.

  • Every room now matches the paper it was drawn from.
  • Where the second sheet sits, measured rather than assumed.
  • A house with a back, for the first time.
  • A sentence telling a model to keep the camera does nothing.
capture.htmlCapture

No photograph of this house exists, so the first one was generated. Then seven video models, a solve gate, and eight reconstructions you can move the camera in.

  • The house does not exist yet, so its first picture was generated.
  • Seven video models from one starting frame. The screening picked the wrong winner.
  • Every generated clip goes through the same solve as real footage.
  • The house was rebuilt from generated video. Load a model and move the camera yourself.
warping.htmlWarping

A solved frame carried to a camera that never existed. Measured by how much of the original survives the trip.

  • Ten small warps land worse than one large one.
angles.htmlMissing angles

Fourteen attempts at the three sides of the house nothing had ever seen, in the order they were tried, failures kept.

  • Every clip starts at the same anchor frame, so the solve never leaves the front.
worlds.htmlWorlds

Nine room plates from before the reconstruction, and the scripted tour.

  • Room plates from before the reconstruction, kept for their labels.
  • One take, ninety-six seconds, every control moving under script.
12Where it stands

Wider coverage cost 1.34 dB, and one view still fails.

Fusing the R4 arc with three new clips put 104 of 104 images into a single model on the first try: 33,572 points at 0.889 px, with lateral camera coverage going from 6.5 to about 13.8 units.

That costs 1.34 dB on the R4 arc in exchange for roughly three times the view volume. R4's worst endpoint got 1.19 dB better, because every new clip sees it.

Fixed

Pull-back driveway

before
fogged by veiling sheets
after
house and wet driveway crisp
obstruction
50.4% → 41.1%

The walk-up view a person actually looks from.

Improved

Yaw extremes

left
neon blobs → real foliage
right
more real palm detail
top-100 share
15.4% / 13.4%

Both edges came back with real detail instead of invented blobs.

Still failing

25° low-eye pitch-down

top-100
71.9% of view weight
scale ratio
256× scene median
anisotropy
2,543

Not a floater problem. No capture yet sees the near ground at a grazing angle.

Next

What that needs

A clip whose camera is low and pitched down while translating, so a driveway approach dolly at −20 to −25°. That is a new control family and a paid decision, so it is not in this build.

full 360 estimate
~18 hops · $11–14

Also on the record: one r5 clip completed and billed at $0.50625 with 175.6 s of inference, and then the result store returned HTTP 500 on all 23 retrieval attempts over 24 minutes. It was left out of the fusion rather than worked around. Request 019fd250-e52d-7713-b0f1-63e242925bc2 is still retryable if that store recovers.

Sealed runs
 

Each one is a manifest with a per-file sha256 index, an authority boundary and a spend record.

Reviewed spaces
51

Across two floors, 33 of them carrying a legible dimension string off the sheet.

Room-perimeter edges
203

Counted from the shipped graph, 117 plus 86. These are room boundaries rather than traced walls; the traced walls are kept in their own manifest as rejected analysis.

NextThe plan

Two phone photographs, and nothing else.