Blueprint
On this pageCurrent buildMissing-angle experiments
360 buildCurrent pipeline

The first 1080p outline-controlled arc exposed identity drift.

An accepted geometry-locked f000 repair and the accepted f020 still anchor a five-second move with a 120-frame animated outline. Ray 3.2 returned 1920 × 1080 video with fluid motion through the expected camera span.

An exact-endpoint composite passed every declared numeric check. Frame-by-frame review still rejected the arc: the roofline, porch depth, window rhythm, side wings and facade proportions change between the accepted endpoints. The 360 dataset remains blocked while stronger identity control is prepared.

Paid result

Rejected

Ray 3.2 returned 120 frames at 1920 × 1080 and 24 fps for $2.16.

Geometry controls

12 accepted

Each five-second segment has its measured animated outline ready for a bounded request.

Appearance anchors

2 admitted

f000 and f020 are accepted at 2048 × 1152. Ten ring anchors still need repair and admission.

Paid evidence

$4.89

Thirteen Flux still/edit requests cost $2.73. The first 1080p Ray arc cost $2.16.

The review composite places the admitted stills at frames 0 and 119. Its interior frames morph the building, so the complete arc remains outside the training set.
Twelve evenly spaced frames from the 1080p Ray arc, showing a continuous camera move around the measured exterior.
The recovered camera path is monotonic, endpoint SSIM measures 0.919 and 0.905, and the intermediate facade still changes identity.
Accepted f000 appearance anchor at the exact first camera of the exterior orbit.
Accepted f000 repair. The measured camera, roof profile, openings and porch pass the sealed local gates.
Accepted f020 appearance anchor at the exact second camera of the exterior orbit.
Accepted f020 anchor. It supplies the shared endpoint for the next bounded arc.
f000 repaird9f815a2e3b302be0401c9f1988e46caafcafbe8e9391d95023865dfa1786482
f020 anchor263924cab129b7c71d51d19e5b92129a17d288fb39b86949bb81834811b36b40
Ray video394d0ff6898f605699b1477f308f74e93688f388697d3866f7a64f7b9ee5e992
Review composite94dea84c91f06f265d0fd5a1c48ef0627cedea9edae103dcca8e8153d274c29c
Open the 360 build in Studio Read the supporting experiments
09Coverage beyond the front facade

Exact cameras extend appearance around the full orbit.

The first fusion holds 104 solved frames, all inside roughly 60° of the front door, because every source clip begins from the same image. It gives the project a readable front arc and leaves the side and rear elevations almost without appearance evidence.

The supporting experiments compare text-directed camera changes, plan-derived reference views, animated geometry control and held-out reconstruction. They establish the control method and show where side and rear appearance still needs new source evidence.

Current finding25 August 2026

Five coherent camera views produced a rejected reconstruction.

Appearance from the strongest front view was propagated through five cameras rendered from the measured house, and the numbers looked great. All four adjacent camera pairs passed the line-coherence check, and a split-safe 20,000-step reconstruction cleared its held-out threshold: mean subject SSIM 0.5100, PSNR 15.0546 dB.

Interactive review exposed elongated rails, shutters, windows and wall planes. Smaller Gaussians reduced the streaks and opened holes in the facade. Camera coherence and aggregate held-out scores passed, while the visible reconstruction failed its product gate.

Five generated house views at exact measured cameras with adjacent-pair line and silhouette comparisons.
The five-view appearance set passed all four adjacent-camera coherence checks.
The five-view Gaussian reconstruction in the live viewer, showing streaked facade detail during adversarial review.
Browser review rejected the splat for hero and product use. The result remains inspectable research evidence.
Inspect the rejected reconstruction
Next controlled experiment

The current run keeps the solved cameras fixed.

Exact-camera stills

Generate close elevation anchors at declared TRAIN and TEST cameras. Each new view must agree with the measured line render before it enters reconstruction.

Matched video study

Run supported video models against the same animated outline, appearance pins, duration and budget. Compare temporal continuity and recovered cameras on equal terms.

Reconstruction sweep

Measure subject-weighted loss, scale and opacity regularization, background treatment and training length on held-out architectural regions, followed by live browser review.

The paid Ray candidate at the top of this page is the first output from this matrix. Side and rear openings still require source evidence from the plans or capture.

09bRejected preflight

Text-only angle requests changed the building.

The 0°, 10° and 20° requests returned nearly identical frames; the 30° request changed the building and the lighting, and a second model family showed the same step response. The $0.44 preflight stopped a planned $5.29 run.

Six generated frames in two rows. The top row, labeled zero degrees, zero degrees again and twenty degrees, is the same wet-driveway front view of the house three times. Bottom left, labeled thirty degrees, is a visibly different building in flat daylight with a garage and no wet driveway. The remaining two frames repeat the test on a second model at zero and ten degrees and return the same picture twice.
The 0° and 20° outputs have cosine similarity 0.9950; repeated identical requests measure 0.9984. The 30° output falls to 0.7947 and changes the house, so the orbit run did not proceed.

Subsequent runs use an animated camera guide rendered from the measured massing. A wide continuous orbit establishes coverage; closer passes follow the facade at a fixed standoff to add detail while retaining the same camera path.

09cPlan-derived reference views

The measured massing supplied the missing structural views.

The July massing contains the first-floor envelope, garages, angled suite wing, double-height great room and foyer, gazebo and covered entry. The same geometry was rendered at four cardinal azimuths. This supplied side and rear structural references with $0.00 in generation cost.

Four renders of the same house massing in a two-by-two grid, labeled front at zero degrees, right at ninety, back at one hundred and eighty, and left at two hundred and seventy. The front shows glazing, a balcony rail, a covered entry and a gazebo. The other three are plain massing with hipped roofs and no windows.
The building bounds measure 47.9 × 8.5 × 30.3; camera orbit and framing use their center. The rear remains bare massing because the source evidence did not contain a reviewed exterior opening schedule.
Plan-derived orbit with 241 frames at 1.5° azimuth increments. Elevation rises from 8° to 26° and returns. Camera schedule: massing-orbit-20260811-r1.

Plan-grounded elements include footprint components, gazebo outline, covered entry, south balcony relationship, angled suite wing, garage recession and the double-height great room and foyer. Wall heights, roof intersections, exterior opening rhythm, materials and terrain remain reconstruction hypotheses.

09dStatic multi-reference test

Static references provided weak geometric control.

Four vendors received the same prompt, one appearance reference and four plan-derived cardinal renders, and the four clips cost $3.31.

Four rows of nine video frames. Each row is one model's clip sampled evenly from first frame to last. Every row shows a large two-story house that changes shape from frame to frame: rooflines, wing counts and window rhythm differ across each row, and none matches the plain gray massing.
Each row contains nine evenly spaced frames. minimax travels furthest and retains the most detail; wan 2.7 shows 9 disturbances where the frame difference jumps past six times the clip’s own median.

All four clips kept the front elevation's appearance while drifting from the 47.9 × 8.5 × 30.3 measured massing. Roof lines, wing count and window rhythm changed across the clips.

The next arm replaced the four static structural references with 241 frames of plan-derived depth and outline control, while keeping one appearance reference. The video guide encodes the intended path frame by frame.

09eAnimated geometry control

The animated outline matched the camera schedule.

One generated frame maps to each of the 241 control frames, preserving the intended 1.5° camera increment. The request cost $1.51.

One continuous orbit circles the building and gains elevation. Appearance comes from the accepted reference still; the path and structural silhouette come from the measured massing.
Six rows of nine video frames. Four reference-only rows each show a different generic two-story suburban house. The two rows given a control video show the long low house from the drawings: one wide, rising to overhead views of its roof planes, one close in along its walls.
Rows one through four received static references and drift from the measured house. Rows five and six received the animated plan orbit and follow its structure. Labels identify each request’s inputs; every clip is sampled at the same nine normalized positions.
A scatter plot in which every one of 241 output frames lands on a straight diagonal line against the control frame it matches, with four pairs of stills below: gray massing renders on top and photorealistic frames of the same building from the same angles beneath them.
Edge-only matching placed each of the 241 output frames along the 241-frame control path. 100% of frames land within tolerance, the advance rate is +1.002, the mean miss is 0.5 frames, and the travel comes out 357°. Peak match 0.582 against 0.306 for a random pairing.

The metric was validated against identity, reversed-order, shuffled-order and uncontrolled-clip cases before it was applied to the controlled output.

The same scatter plot for a different clip, with matched points scattered in a shallow cloud away from the diagonal, and a note identifying the clip as off the control path.
wan 2.7 placed 0% of frames on the control diagonal, with an advance rate of +0.254, mean offset 92 frames, and only 34 of 241 control frames matched by anything at all. Its scattered matches span 354°, demonstrating that angular span alone does not measure path coverage.

The controlled clip contains no cuts and maintains continuous travel. Its largest frame-to-frame difference is 1.11× its median, compared with 8.59× and nine disturbances for the strongest uncontrolled clip. The first and last frames retain the same building and planting layout.

The brick corner tower is absent because it is absent from the massing. Geometry control governs structural form, while the appearance reference supplies material and lighting. Haze increases near the top of the elevation ramp, where the reference provides the least visual evidence.

BackDepthNextStudio

Direct the environment from one shared canvas.