| name | scene-to-layout |
| description | Turn a room photo, scan, or point cloud into a structured layout — walls, openings, furniture as typed objects with dimensions — using SpatialLM-class models, ready for blueprint-validate and re-design. Trigger: "here's a photo of my room", "measure this space", "redesign from this picture".
|
| kind | skill |
| license | CC-BY-4.0 |
| runtimes | ["claude-code","codex","hermes"] |
scene-to-layout
When to use
The input is a picture (the repo's second input modality) rather than text —
an existing room/site that must become structured data before redesign.
What it does
- Route by input: photo → monocular layout estimation; scan/point-cloud →
SpatialLM (https://github.com/manycore-research/SpatialLM); floor-plan
image → wall/opening extraction.
- Emit a structured layout JSON: walls (start/end/height/thickness), openings
(type/position/size), furniture (class/bbox/orientation), declared units.
- State per-element confidence; dimensions estimated from a single photo are
marked ESTIMATED and need one user-confirmed reference measurement.
- Hand off to blueprint-validate (audit) or brief-to-blueprint (redesign).
Example
"Photo of a 4x5m living room + 'the door is 80cm' reference" → layout JSON with
scaled dimensions → "now add a reading corner with daylight" → redesign loop.
Verification (eval-with-teeth)
Output JSON must round-trip: rendered back to a 2D plan, wall topology is
closed (rooms are bounded polygons) and openings lie on walls. An unclosed
room polygon is a failed extraction — report it, don't patch it silently.
Safety
Never claim measured accuracy from an unreferenced photo; renovation decisions
need a tape measure.
Cross-runtime
Model access varies by runtime; degrade to "ask user for measurements" when no
vision/SpatialLM backend is available (skip-not-fail).
