Accepted source · spatial candidate
Exterior arrival
The exterior keeps the strongest admitted appearance while orientation, held-out geometry, and threshold alignment remain under review.
Blueprint Studio
Turn plans, images, and prompts into living 3D worlds.
Plans, source images, generated views, reconstructed scenes, actors, cameras, and world state live in one persistent editor. Build a place, direct it, and enter it from the same canvas.
- World
- 10 registered cells
- Simulation
- 5 persistent actors
- Portal authority
- 4 blocked · alignment pending
- Representations
- Source · depth · points · splat
Connected-space proof
Arrival, threshold, interior.
The atlas is visible now. Camera focus is available across cells; actor transfer stays blocked until correspondence, visibility, scale, ground, and continuity checks pass.
Capture gap · threshold blocked
Front-door foyer
The foyer has a persistent cell identity and presentation evidence. It still needs a coherent threshold capture before portal alignment can be admitted.
V22 playable authority
Great room
The great room retains its measured local floor, two distinct animated actors, directly editable routes, point field, Gaussian splat, and source-aware render controls.
Production method
Keep the appearance. Test the geometry.
Each run starts from registered source evidence, then moves through depth, camera recovery, reconstruction, and browser review. Structure, held-out views, ground, memory cost, and portal conditions are checked separately because a convincing image can still describe the wrong space.
Failed runs stay attached to their inputs, timestamps, hashes, metrics, and known failure modes. The current catalog contains 40 unique splat payloads across 50 physical files, including the weak experiments that explain why the current method changed.
Exterior reconstruction
The route back to a coherent house.
The strongest stills were already good enough for presentation. The difficult part was turning them into multiple views of the same structure without asking a generative model to redesign the house between frames.
Appearance
Admitted high-resolution stills
The strongest FLUX stills established the appearance target before reconstruction.
Camera evidence
Measured outline at frame 000
A traced massing model supplied exact camera motion and kept generated views tied to the same house.
Cross-view review
Split-safe camera coherence
Adjacent views were checked against the known camera geometry before any splat training was allowed.
Current candidate
Observed-surface source assignment
The current bounded candidate assigns the best observed TRAIN view to each traced facade point.
Surface Bake R2 remains the cleanest wider-arc presentation of observed facade pixels. The five-view Brush reconstruction has now completed and was rejected: its narrow numeric pass did not survive browser inspection. The accepted six-degree micro-parallax result therefore remains the current exterior viewer candidate.
What the runs established
Useful failures changed the method.
retained failure evidence
Independent views drifted
Adding more generated stills increased coverage but also changed chimneys, siding, window layouts, and the depth of side walls. The All-60 result remains an editable working base, not a final house.
bounded use
Provider video had a narrower role
Ray, Kling, Seedance, and related runs supplied useful motion and appearance evidence. They were rejected as metric camera authority whenever walls bent, motion stepped, or the control image leaked into the rendered house.
bounded presentation
Observed surfaces stayed tied to known cameras
Surface Bake R2 assigns each traced facade point to its best visible TRAIN observation. It produces a cleaner upright presentation across the measured 29.8755-degree arc while leaving unobserved walls absent.
rejected after browser review
The Brush metric pass was visually rejected
The five-view 20k refit reached 0.5100 mean subject SSIM and 15.0546 dB PSNR, barely clearing its numeric thresholds. Browser review still found facade ghosting, elongated streaks, missing architectural detail, spikes, and floaters, so the result remains negative evidence.
Local project measurements
Fast previews, slower authority.
These are single-run measurements from the RTX 5080 workstation, not general model benchmarks. The measured 28-frame pose solve plus Brush training took about 7 minutes 49 seconds before extraction, inspection, and publication.
- Depth preview
- 1.576 s DA3 small on 34 known-camera frames
- Pose solve
- 3.030 min COLMAP on the 28-frame fusion-KR run
- Radiance field
- 4 min 47 s Brush 30k on 22 TRAIN and 6 TEST views
- Current facade bake
- 2.42 M observed points at 1.75 cm spacing
Current product boundary
One world, one viewport, honest handoffs.
Exterior, foyer, and Great Room keep stable cell identities in the same world. The Great Room is the accepted playable reconstruction with distinct animated actors and directly editable persistent routes. Exterior and foyer remain inspectable while their portal transforms stay blocked.
Camera, Render, Edit, Evidence, Generate, and Run are workspaces around one React Three Fiber canvas and one Gaussian renderer. Selecting a result changes the shared viewport; it does not launch another viewer.