Where Generated Floor Plans Stop: Calibrated Sample-Level Checks Against a Built Baseline
Organizations: Skidmore, Owings & Merrill (SOM), Chicago
Abstract
Floor plan generators are judged mainly by Compatibility, a graph edit distance to the requested adjacencies that is near saturation and blind to room shape. Our sample-level protocol calibrates each threshold no standard states at a declared false-rejection rate on real samples and grades them with the same checks. It reports each sample at its first failure, counts samples without a verdict apart, and attributes a stop to the generator or to its representation where a control exists. We apply it, with a production layout solver's checks, to four published generators against two built corpora, RPLAN and Modified Swiss Dwellings (MSD). A squareness check set per room type refuses 5 per cent of RPLAN apartments and 33.7 per cent of House-GAN++ plans. The gap survives resolution, tracing and orientation controls. Pooled over rooms, a Wasserstein distance on squareness orders the generators nearly as the check does but separates House-GAN++ from the built corpora by far less. Taken per plan on the check's worst-room margin, it separates them about fivefold to sevenfold. Within the GAN generators, Compatibility barely predicts which plans pass. We release the checker, which reproduces the solver's first failing check on 3,998 held-out RPLAN plans.
Figures & tables
| Check | What it decides | Where the threshold comes from | On public datasets |
|---|---|---|---|
| 1 Room outlines roomOutline | Geometry alone, on a plan ingest has accepted | No request-wide constant and no per-room requirement | Runs; no decision recorded |
| 2 Room compactness roomCompactness | The least compact non-hallway room against a Polsby-Popper threshold ( Polsby & Popper, 1991 ) | Fitted so the check alone refuses 5 per cent of real plans of each kind: 0.365 RPLAN apartments, 0.3237 MSD apartments, 0.2406 MSD floors; no external validation | Runs, corpus-fitted |
| 3 Room proportions roomProportions | Each room’s squareness, short side over long of its unrotated bounding box, against a threshold per room type | One quantile of each room type’s squareness, chosen so the check alone refuses 5 per cent of real plans of each kind; no external validation | Runs, corpus-fitted |
| 4 Room appendices roomAppendices | Whether a protruding sub-area counts as usable space, once the room’s walls form a closed loop meeting at right angles | One of the two conditions in each of the policy’s two tests derived; the precondition’s tolerance derived | Policy disabled; precondition stops the plan without a verdict |
| 5 Suite container suiteContainer | Whether a suite contains the rooms it declares | None needed: the data carry no suites | Not transferable |
| 6 Final adjacency constraints finalAdjacency | Whether the final geometry still meets the requested adjacency program; a request with no program leaves it nothing to test | None needed: the data state no adjacency program | Not transferable; reported as skipped |
| Source | Graded | Passed | Compactness | Proportions | Right-angle precondition | Ingest refusal | Other |
|---|---|---|---|---|---|---|---|
| RPLAN ground truth | 4,000 | 91.3 | 5.0 | 3.7 | 0.0 | 0.0 | 0.1 |
| MSD ground truth, per apartment | 14,324 | 70.3 | 4.8 | 3.9 | 20.1 | 0.8 | 0.0 |
| MSD ground truth, per floor | 4,167 | 57.8 | 5.4 | 3.7 | 28.8 | 0.9 | 3.5 |
| House-GAN++ | 4,000 | 58.2 | 2.5 | 28.3 | 0.0 | 11.1 | 0.0 |
| House-GAN | 4,000 | 48.7 | 20.6 | 19.6 | 0.0 | 11.1 | 0.0 |
| HouseDiffusion | 3,991 | 1.1 | 21.1 | 1.1 | 2.4 | 70.4 | 3.9 |
| Generator | Where it stops, per cent of graded plans | Attributed to | What decides it |
|---|---|---|---|
| House-GAN++ | Proportion check, 28.3 | The generator | RPLAN in House-GAN++’s own training rendering is refused by the check alone at 5.8 per cent, against 33.7 per cent of its own plans (Section 4.2.2) |
| House-GAN | Compactness, 20.6, then the proportion check, 19.6 | Proportions: its representation; compactness: not settled | RPLAN rendered as its training targets is refused by the proportion check alone at 34.2 per cent, against 32.5 per cent of its own plans; by compactness alone at 10.2 per cent against 23.2 per cent of its own, controlled at the unrefitted threshold only (Section 4.2.2; Supplementary analyses, Section S3) |
| HouseDiffusion | Ingest, 70.4 | Not settled by a control. The overlaps are substantive: in the 62.7 per cent of plans where two rooms claim the same floor, the median claimed twice is 1.351 square metres, and 29.8 per cent are refused with no overlap rule at all. This points to the generator, although no control settles it | Specification, section 9.2; it contributes no evidence to the margin |
| GSDiff | Right-angle precondition, 83.3, without a verdict | Not settled. Its segments run off the axes by more than one raster cell, and rectifying at one cell recovers 32 of 2,425 stops. RPLAN in the same pass stops there at 0.0 per cent, but no real corpus is graded as the faces of a wall graph | Restoring the one-cell rectify threshold recovers 32 of its 2,425 stops (Section 3.2); with its merged living, dining and foyer class typed, the proportion check alone refuses 7.7 per cent (Supplementary analyses, Section S4) |
| Source | Plans the checker’s ingest admits | Proportion check alone refuses (%) | Compactness check alone refuses (%) |
|---|---|---|---|
| House-GAN++ | 3,557 | 33.7 | 2.8 |
| House-GAN | 3,556 | 32.5 | 23.2 |
| HouseDiffusion | 1,189 | 19.8 | 82.2 |
| GSDiff | 2,820 | 6.9 | 3.9 |
| RPLAN ground truth | 3,998 | 5.0 | 5.0 |
| MSD ground truth, per apartment | 14,206 | 4.9 | 4.8 |
| Control | From RPLAN, thresholds unchanged | From RPLAN, thresholds refitted | From the MSD apartments, thresholds unchanged |
|---|---|---|---|
| None, geometry as measured | 28.8 | 28.9 | |
| Centre sampling at 0.28125 m, 16 grid phases | 28.0 (27.7 to 28.4) | 26.5 (25.9 to 27.0) | |
| Majority coverage at 0.28125 m, 16 grid phases for RPLAN, 4 for MSD | 28.5 (28.2 to 28.7) | 27.7 (26.2 to 30.6) | 28.7 |
| House-GAN++’s own training rendering at its frame | 28.0 | 24.8 | |
| Opening that deletes every part up to two pixels wide | 28.4 | 28.1 | |
| Minimum-area rectangle in place of the unrotated box | 28.7 |
| Condition | Graded, G | Admitted, A | Reached a verdict, V |
|---|---|---|---|
| Reported run, each corpus on its own fit | +12.1 | +5.5 | +23.5 |
| Same, 95 per cent interval with whole MSD buildings resampled | 9.7 to 14.6 | 3.0 to 7.9 | 21.5 to 25.6 |
| Term: proportion check | +24.4 | ||
| Term: ingest refusal | +10.3 | ||
| Term: right-angle precondition | -20.1 | ||
| Term: compactness | -2.3 |
| Constant | What it governs | Anchor | Value |
|---|---|---|---|
| joinTol | Endpoint distance at which wall segments chain | Half an RPLAN cell | 0.03515625 |
| rectifyThreshold | Deviation from horizontal or vertical that the construction step snaps straight | One RPLAN cell; superseded by decision 5 of Appendix B | 0.0703125 derived; 0 in force |
| abstol | Short-segment and duplicate-vertex removal; the identity quantisation | One millimetre, about one seventieth of an RPLAN cell | 0.001 |
| angtol , rad | Collinearity angle when simplifying a polygon | One cell of staircase over one door width, atan(0.0703125 / 0.813) | 0.0862706 |
| localtol | Whether a span lies on a boundary | One RPLAN cell | 0.0703125 |
| minPassingPPScore | Polsby-Popper compactness threshold | Corpus-fitted, no external validation: no statutory or peer-reviewed threshold; a square scores 0.7854, a 2:1 rectangle 0.6981, a 3:1 rectangle 0.5890 | Plan-level fit at a 5 per cent false rejection: 0.365 RPLAN apartments, 0.3237 MSD apartments, 0.2406 MSD floors (reported run) |
| # | Decision | Specification section |
|---|---|---|
| 1 | MSD floors are divided into rooms from their exact geometry rather than from a grid-snapped conversion | 5.6 |
| 2 | The exact-geometry conversion welds features up to 0.40 m, the widest wall the corpus draws | 5.6 |
| 3 | A plan whose rooms do not tile the plan’s outline is refused at ingest and counted apart | 6 |
| 4 | An open gap between rooms that do not touch is closed by the same rule as an enclosed void, for every source | 6.3 |
| 5 | The rectify threshold is zero, decided when one RPLAN cell refused 65 per cent of House-GAN++ plans inside the straightening step | 8.8 |
| 6 | The compactness threshold is fitted per kind so that the check alone refuses 5 per cent of real plans | 8.2 |
| Rule at 0.5625 m | From RPLAN, thresholds unchanged | From RPLAN, thresholds refitted | From the MSD apartments, thresholds unchanged | From the MSD apartments, thresholds refitted |
|---|---|---|---|---|
| Centre sampling, 64 grid phases | 25.0 (24.5 to 25.5) | 21.6 (19.6 to 24.9) | ||
| Majority coverage, 64 grid phases for RPLAN, 4 for MSD | 26.0 (24.9 to 26.8) | 23.0 (19.8 to 26.2) | 25.1 | 16.0 |
| Proportion threshold | Real RPLAN plans refused (%) | Separation of House-GAN++ from RPLAN (points) |
|---|---|---|
| Fitted per room type, the reported run | 5.0 | +29.0 |
| Single, 1:3 | 28.7 | +7.7 |
| Single, 1:2.5 | 50.5 | +9.0 |
| Single, 1:2 | 79.7 | -0.5 |
| Single, the bedroom cap of 0.60 | 95.0 | -2.0 |
| Single, 1:1.5 | 98.4 | -1.6 |
| Ladder | Graded | Admitted | Reached a verdict |
|---|---|---|---|
| The checker’s fitted ladder | +12.2 | +5.6 | +23.6 |
| Width rule alone | +22.3, RPLAN over House-GAN++ | +6.6 to +11.6, RPLAN over HouseDiffusion | |
| Width rule in place of the proportion threshold | +1.7 (+0.1 to +3.3) | -6.6 (-8.1 to -5.0) | +10.6 (+9.1 to +12.3), RPLAN over House-GAN++ |
| Same, scales moved 5 per cent either way or the width threshold one frame step | -4.8 to +8.7 | -13.9 to +1.3 | +3.3 to +18.5 |
| Checks 1 and 4, then the width rule | +3.0 | -5.5 | +2.9, RPLAN over HouseDiffusion |
| Same, scales moved 5 per cent either way or the width threshold one frame step | -3.8 to +10.3 |
| Generator | Population | Passing | Failing | AUC, pooled | AUC, within bucket |
|---|---|---|---|---|---|
| House-GAN | G | 1,948 | 2,052 | 0.504 (0.487 to 0.521) | 0.451 (0.433 to 0.467) |
| House-GAN | A | 1,948 | 1,607 | 0.487 (0.468 to 0.506) | 0.432 (0.413 to 0.452) |
| House-GAN++ | G | 2,327 | 1,673 | 0.531 (0.513 to 0.548) | 0.519 (0.501 to 0.536) |
| House-GAN++ | A | 2,327 | 1,230 | 0.472 (0.453 to 0.490) | 0.477 (0.458 to 0.496) |
| HouseDiffusion | G | 42 | 3,949 | 0.777 (0.724 to 0.826) | 0.666 (0.599 to 0.729) |
| HouseDiffusion | A | 42 | 982 | 0.645 (0.574 to 0.712) | 0.607 (0.532 to 0.681) |
| Source | Labels and the type each reaches | Labels that fall to Unset |
|---|---|---|
| RPLAN | livingroom, LivingRoom; masterroom, PrimaryBedroom; childroom and secondroom, Bedroom; guestroom, GuestRoom; kitchen; bathroom; diningroom, DiningRoom; balcony; entrance, Foyer; storage and walkin_closet, Closet | studyroom |
| MSD | bedroom; livingroom, LivingRoom; kitchen; dining, DiningRoom; corridor, Hallway; storeroom, Closet; bathroom; balcony | stairs, any unknown code |
| House-GAN++, HouseDiffusion | living_room, LivingRoom; bedroom, from RPLAN’s primary, child, second and guest bedrooms, Bedroom; kitchen; bathroom; balcony; entrance, Foyer; dining room, DiningRoom; storage, from RPLAN’s storage and walk-in closet, Closet | study room, unknown |
| House-GAN | living_room, LivingRoom; bedroom, from RPLAN’s bedrooms and study room, Bedroom; kitchen; bathroom; balcony; corridor, from RPLAN’s entrance, Hallway | closet, dining_room, missing, laundry_room |
| GSDiff | bedroom; storage, Closet; kitchen; bathroom; balcony | living/dining/foyer; exterior is excluded before export |