Blueprint | Reconstruction workspace

04Starting image

Build from evidence.

The first capture experiment began from a generated front elevation. It predates the measured structure and exact-camera FLUX studies documented in the Case Study.

A traced plan and style brief produced this image. The evaluation loop compares candidate motion, solves cameras, trains registered views and scores held-out views. Its narrow front coverage defined the need for exact multi-angle references.

The generated front elevation: a two-story
cream lap-sided house with a brick bay tower on the right, a glazed sun-room wing on the left, hipped
standing-seam metal roofs, warm light in the windows, wet brick walk and driveway, live oaks and
palms around it. — Legacy generated front elevation. The source PNG has no camera EXIF and carries its generation provenance. Treat it as an appearance reference; no camera ever measured this view of the property.

The source image shows 2 stories, 4 porch columns, 3 balcony bays, a five-sided brick bay tower on the right and a single-story glazed sun-room on the left. The asymmetry distinguishes the intended orientation from mirrored or structurally changed frames.

05Model comparison

Seven video models started from the same frame.

All seven ran from the identical generated still with the same prompt and negative prompt, then faced a local screening proxy: feature continuity, camera motion, source retention, cut safety. Thirteen runs across two scenes, $8.44 actual against a $9.00 budget.

The proxy ranked Veo first at 90.9 and Kling fifth at 85.3. Structure from motion reversed that order: Kling reconstructed and Veo failed. The proxy filters obvious visual failures. Camera recovery determines whether a clip contains usable multi-view geometry.

Screening formula: 35% feature continuity, 25% camera motion, 25% source retention and 15% cut safety over SIFT correspondences with RANSAC fundamental geometry. Persistent 2D texture can score without coherent multi-view structure, so every passing clip also receives a camera solve.

06Camera recovery

Camera recovery gates every clip before training.

Each clip goes to structure from motion under pinned intrinsics, and the score counts how many of 210 image pairs return a geometrically calibrated relationship. A clip with no stable camera solution stops before any paid generation or reconstruction training.

Camera recovery and visual quality are separate review gates. The high-scoring a150 guide failed on roofline and foliage artifacts. The pb100 guide failed local camera recovery and stopped before generation.

Camera guides · solved locally before generation

Superseded

t025 · dolly parallax

calibrated pairs: 124 / 210
family: r6

The first guide that produced a solvable generation at all.

Selected arc

a100 · 1.00u truck, 18° sweep

calibrated pairs: 189 / 210
points: 8,275

The middle arc of the r7 family, selected for the next generation.

Rejected: visual artifacts

a150 · 1.50u truck, 26° sweep

calibrated pairs: 198 / 210
solve: 21/21 · 7,714 pts · 0.847 px
why out: roofline bend, foliage smear, 5 watermark pairs

Highest guide solve score; rejected for roofline bending, foliage smear and five watermark pairs.

Best parallax

l100 · lateral dolly 1.00u

guide pairs: 190 / 210
clip pairs: 206 / 210
points: 8,060 · 0.765 px

The guide calibrated 190 of 210 pairs; its generated clip calibrated 206.

Rejected: camera solve

pb100 · pure pull-back

two-arm solve: no sparse model
fallback: 8 / 21 · 122 pts
saved: $0.50625

Pure pull-back motion supplied insufficient lateral parallax for mapper initialization, so the candidate stopped after the local solve.

Local preflight result

Six candidates were rendered and solved locally at no cost. Five passed at 21/21 registered, 0.73 to 0.82 px, zero watermark-configuration pairs. One failed outright. Paid generation proceeded for the five passing guides.

candidates: 6
passed: 5
paid on failures: $0.00

Generation sequence

Rejected

R2

degenerate pairs: 210 / 210

Visual review passed, but every pairwise solve collapsed into one degenerate camera configuration.

First pass

R3

calibrated pairs: 132 / 210
points: 8,576

The first generated clip that reconstructed at all.

Arc baseline

R4

calibrated pairs: 184 / 210
points: 10,480
holdout: 27.78 dB

Sharpest within its solved arc; unsupported views collapse outside that range.

First corner gain

Orbit hop 1

joint solve: 128 / 128
coverage: 35.4° → 53.7°
calibrated: 204 / 210

Strongest parallax in the project, and the first view past the corner.

06bHoldout PSNR

A controlled baseline anchors reconstruction comparisons.

A controlled re-derivation of the strongest prior generated clip measured 30.34 dB on its held-out split. Later candidates use this value only when the evaluation protocol matches.

The strongest generated candidate in this study measured 27.78 dB. Direct viewer review remains a separate gate because a reconstruction can score well inside its source arc and fail as soon as the camera leaves supported coverage.

07Blueprint Studio

Four reviewed states open in one spatial environment.

Choose the accepted front appearance, the geometry-propagated surface study, the held full-turn research base or the playable great room. Each state opens in the canonical Studio with its camera, review scope and evidence attached.

Camera, Render, Edit and Evidence operate around the same world while the visible representation changes.

Choose a Studio state

Open selected state in Blueprint Studio

Current bounded exterior appearance, reviewed across a six-degree front arc.

Measured viewing range for each reconstruction.

Viewing bounds are based on direct review of 52 renders: four models at thirteen azimuths in 10° steps, each marked clear, degraded or gone. Two automated alternatives were rejected. Sharpness rises when Gaussians stretch and foliage fragments. Solved camera positions describe capture locations; direct review establishes the readable range.

A contact sheet of 52 rendered frames: four reconstructions down the side, thirteen azimuths from minus 60 to plus 60 degrees across the top, each frame carrying a colored rail marked clear, degraded or gone. The middle of every row is the house; both ends are colored streaks. — Direct-review frames at their evaluation size. Scroll sideways. Along each row the house remains readable through the middle and then breaks up, with earlier failure on the left: turning left reaches an unresolved foreground oak and palm, while turning right opens onto the driveway the clips did cover. Every model reaches farther right than left, requiring asymmetric viewer limits.

Four reconstructions on one azimuth ruler. For each, small ticks mark the solved camera positions, a filled block marks the arc that still reads as the house, and an outline marks how far the viewer lets the camera go. — Ticks show solved camera positions, the filled band shows direct visual acceptance and the outline shows the viewer clamp. Capture positions and readable viewing range remain separate measures. **kling-rederive-30k** now opens across **30°**, compared with its **5.9°** solved-camera span.

Readable range saturated as capture widened.

Across the four models, readable degrees gained per camera degree fall away steadily: 5.06× readable degrees per camera degree, then 1.75×, then 1.01×, then 0.83×. A 10× spread in capture width bought 20° of additional readable view, and the two widest captures return their own solved arcs and little more. Beyond a point, the win has to come from source appearance and reconstruction quality.

Readable arc plotted against camera arc for four reconstructions, with the one-to-one line drawn. The points rise from 30 degrees readable at 6 degrees of camera to 50 degrees readable at 60 degrees, crossing from far above the line to below it. — Above the dashed line, readable range exceeds the camera arc; below it, the reconstruction returns less range than the captured cameras. Data: splat-arc-bands.json.

Viewer-range provenance

The current bands were generated from the sealed viewer-arc-20260811-r2 manifest. The replaced instrument and its full-resolution frames remain in the archive as rejected range evidence.

Transparent training masks were read as empty space.

Feeding generated frames into a reconstruction costs −1.43 dB against a baseline built from solved frames alone, measured on held-out views. A warped frame mixes resampled source pixels with invented disocclusions, and the warp records that provenance per pixel, which suggested marking invented regions transparent so the trainer could ignore them. A controlled test punched a rectangle to transparent in half the training views to see which way the trainer reads transparency; the other views and all training settings stayed untouched.

Three crops of the same balcony:
the source frame, the model trained on intact images, and the model where half the training views had a
rectangle punched to transparent. Inside the outlined rectangle the third crop is pulled dark, the lit
window dims and the white rail grays, while everything outside it matches. — Transparency was interpreted as empty space. Inside the masked rectangle, held-out quality fell by **5.53 dB**; outside it, the two arms differed by **0.11 dB**. The effect ends at the rectangle boundary, and setting alpha-loss weight to zero leaves it active.

So transparency means empty space to this trainer, and masking real disocclusions that way deletes geometry other views actually support. Two short training runs completed in about four minutes. The result ruled out a larger per-pixel masking experiment. Whole-frame selection by invented-pixel fraction remains untested.

Alpha-semantics provenance

The controlled result is preserved as refuted in alpha-semantics-20260811-r1. It stopped the larger per-pixel masking experiment before additional training.

Prepare a spatial reconstruction.