Teaching a camera to remember the room well enough to set the shot.
The research twin of FocalaneCapture, teaching it to know a creator's filming spaces well enough for an agent to set the shot, not merely see the room.
Creators film in the same few rooms, yet every session starts from scratch. The light has drifted, the tripod sits somewhere new, and the last good setup lives only in memory. A camera frame shows what is visible now; it cannot say what changed or what would help this video. In Christian's words, FocalaneSpatialLab is "where we teach FocalaneCapture to know a creator's familiar filming spaces well enough for an agent to set the shot, not merely see the room."
Scan a recurring room once and remember its geometry and its gear. On filming day, recognize the room, show what moved, and join that memory with the video's intent, the creator's history, and fresh camera views to propose where the camera, lights, and creator belong. The real camera then checks every suggestion. A later idea is a pre-shoot preview: the creator's cutout placed in that room to compare spots, framing, and light.
The bet is that measured geometry, not a guess from one frame, is what lets an agent give an honest instruction. Today, the lab's app on an iPhone 16 Pro has saved a room and relocalized into it in 3 of 3 fresh-process trials, and its what-moved comparison works on synthetic test data. Recognizing the real light, tripod, and chair is in progress. No agent sets a shot yet and no preview exists; the lab only proposes, and FocalaneCapture decides.
How it works
Sense and remember
Recognize and compare
Propose
FocalaneCapture decides
A separate research app and a bounded Mac worker turn capture-owned spatial evidence into setup proposals. FocalaneCapture accepts, rejects, or ignores each proposal; the lab owns no production camera, accepted setup, or product state.
Where it stands today
On an iPhone 16 Pro, the lab's separate research app saved a room reference with depth and mesh, relocalized into it in 3 of 3 fresh-process trials, passed a bounded RoomPlan scan, and completed a cooled ten-minute run. On the Mac side, the contracts, the local worker, and a setup comparison that labels items as moved, unchanged, missing, new, or ambiguous pass on synthetic fixtures. Recognizing the real key light, tripod, and chair is still in progress, and a performance-stabilization build is installed from a local, unpushed branch and not yet run attended; no performance improvement is claimed. There is no shot-setting agent and no creator preview yet: the lab emits proposals only, and FocalaneCapture has integrated none of them.
Working now
Saved-room relocalization on iPhone 16 Pro (3 of 3 fresh-process trials)
Depth, mesh, and bounded RoomPlan room capture
Setup comparison that explains moved, missing, new, and ambiguous items (synthetic fixtures)
Proposal-only evidence contracts and a bounded local Mac worker
Thermal and lifecycle policy with atomic evidence storage (automated tests; device run pending)
Not there yet
Recognizing the real key light, tripod, and chair is in progress, and a physical setup comparison is not yet proven.
The latest performance build is installed; its attended run is pending.
How it got here
Founded as an isolated research repository, from an earlier conversation about a lasting model of a filming space: where things are, what they are, which one is which, and an agent reasoning about the shot on top
iPhone 16 Pro gates: room reference, 3 of 3 relocalizations, a RoomPlan scan, and a cooled ten-minute run after a failed charging-condition run
Research paused partway through recognizing the physical filming equipment
Named FocalaneSpatialLab; the set-the-shot thesis accepted; a performance-stabilization build implemented
FocalaneCapture reviewed the lab's contracts as review input, not an integration dependency
What comes next
Next step
Attended iPhone 16 Pro run of the stabilization build, then resume recognizing the three filming anchors.
Later ideas, not current features
Attended iPhone 16 Pro run of the stabilization build: fresh reference, three relocalizations, traces, and a ten-minute heat check
Finish recognizing the three filming anchors across held-out lighting, viewpoint, occlusion, and moved-object conditions
An identity-backed physical setup comparison, emitted as proposals
A contract-only review with FocalaneCapture, only after separate authorization
What I checked, and what stays bounded
Dated evidence
iPhone 16 Pro relocalization 3 of 3, RoomPlan scan, and a cooled ten-minute rundevice-test
Performance-stabilization build signed and installed from a local branch; attended run pendingbuild
Boundaries it keeps
No agent sets, moves, or changes a shot today; the lab only proposes, and FocalaneCapture keeps camera and setup authority
Not integrated into FocalaneCapture and not a released product
Recognition of the real equipment and a physical setup comparison are not yet proven; the what-moved result is proven on synthetic fixtures
The lab has no subject segmentation, compositing, or lighting preview; segmentation elsewhere in the Focalane family is used only on recorded takes
A composite preview would be a creative preview, not a promise that the real room will look that way
No thermal cause or before-and-after performance percentage is claimed
Illustrative diagrams are not measured geometry
My role
Directed the research and its authority boundary, ran the attended iPhone 16 Pro gates, and had the runtime stabilization implemented and tested.