A network trained to recover the walls, openings, and rooms of a rasterized floorplan can produce its output in two ways: by emitting the geometry as an autoregressive coordinate sequence, or by detecting it on dense junction and centerline heatmaps and assembling a graph. We compare the two readouts on the same trained network. On real scans (CubiCasa5K) detection is better on every wall measure (+2.7 wall F1 at tolerance 0.05, +5.1 at 0.015; paired bootstrap intervals exclude zero), and reading an opening heatmap the decoder never used raises opening F1 by 2.6x without retraining. Within real scans the readout's advantage grows with plan size and reverses on small plans; on clean vector renders sequence decoding is better by 5 to 8 points where its training covered the render style, while under full domain shift the readout, given calibrated thresholds, stays ahead; neither ink density nor plan size explains the reversal. With matched data and recipe, a room-centric system with a reconciliation step and a wall-first sequence model reach comparable wall quality, so the output representation matters less than is usually assumed. A prior from the other family helps at the output but not at the input: deterministic fusion of the two outputs raises wall F1 by 7 points, whereas conditioning one model on the other's output gives no gain in three forms, including two ground-truth-content controls. We also provide an edit-cost metric that scores a draft by the human work needed to correct it, corrected CubiCasa5K annotations, and ResPlan-FP, a CC BY 4.0 benchmark of 16,998 plans with frozen splits and three baseline tracks. Code, the benchmark, and the corrected annotations are available at https://github.com/Cyprinus12138/fpvec-lab
Figures & tables
Figure 1: The wall-first network and its two readouts. One encoder feeds a grammar-masked sequence decoder (emit) and a dense branch of three heatmaps; the graph readout (detect) assembles walls from the junction and centerline heatmaps without a trained parameter. The room-centric system is a separate network (Raster2Seq) followed by a reconciliation step.
1. Nodes. V← local maxima of HJ with HJ≥θJ (non-maximum suppression), refined by O .
2. Candidate edges. For every pair (u,v)∈V2 : sample the chord uv at n interior points (trimming the ends); accept if the fraction of samples with HC≥θC is at least κ , ∥u−v∥≥ℓmin , and no third node lies within ϵ of the chord (composite chords are rejected; their sub-edges carry them).
3. Chain. At each node, continue an edge into its single best anti-parallel partner within angle α ; walk chains in both directions to obtain maximal near-straight walls. Disabled in every frozen configuration: the ground-truth walls of all three domains are split at junctions, and chaining costs 22 pp on CubiCasa5K validation ( Appendix E ).
Output: walls as chained segments with shared node instances, hence closed by construction; rooms as faces of the planar graph.
Table 1: Graph readout from dense heatmaps to a wall graph. HJ and HC are the network’s junction and centerline heatmaps; no parameter is trained.
Table 2: Same trained network (walls-only arm) with the wall readout swapped, CubiCasa5K official test ( n=296 ). Detection is ahead of autoregressive decoding on every wall measure. Rooms are the faces of the wall graph; the readout emits no openings. The room-centric system is shown for reference ( n=287 ). Per-system 95% bootstrap half-widths on wall F1 are 0.012 – 0.018 .
Figure 2: Qualitative results. Columns: input; sequence decode (walls-only arm); graph readout of the same network; room-centric system after reconciliation; fusion (base in purple, ink-gated donor walls in red); ground truth. Rows 1–4 are CubiCasa5K official-test plans (a typical plan, a plan where the readout is far ahead of decoding, a plan where fusion repairs a weak room-centric output, and a failure case on which every system is poor); rows 5–6 are ResPlan-FP zero-shot plans, one on which the room-centric arm emits no rooms and the fused system falls back to the readout walls. The ResPlan readout panels use the frozen CubiCasa5K thresholds (the zero-shot protocol of Table 10 ), not the in-domain-calibrated variant. Wall F1 at t=0.05 in each title.
domain
n
dark-pixel frac.
GT walls
seq. decode
graph readout
readout − decode
CubiCasa5K (scans)
296
0.078
28.9
0.790
0.818
+2.7[1.4,4.0]
open-plan (clean synth)
30
0.112
35.0
0.908
0.825
−8.2[−12.3,−4.7]
ResPlan, zero-shot (frozen thr.)
1000
0.094
31.0
0.640
0.621
−1.9[−2.7,−1.2]
ResPlan, zero-shot (calibrated thr.)
1000
0.094
31.0
0.640
0.677
+3.7[+3.0,+4.4]
ResPlan, finetuned
1000
0.094
31.0
0.968
0.915
−5.3[−5.5,−5.0]
Table 3: Emit versus detect across domains (wall [email protected]) with the mean dark-pixel fraction of the 256 px input and the paired bootstrap 95% interval of readout minus decode. The walls-only arm is used everywhere except the finetuned ResPlan row. Readout thresholds: CubiCasa5K frozen; open-plan calibrated on an in-domain set ( 0.819 with the frozen CubiCasa5K thresholds); finetuned ResPlan calibrated on ResPlan validation ( 0.914 with the frozen CubiCasa5K thresholds); the two zero-shot rows show the frozen thresholds carried over and the effect of calibrating only the thresholds on ResPlan validation (no training). The winner is not monotone in ink density or in plan size.
system
wall F1
wall P
wall R
rooms
open
edit
r2s_cc5k (rooms+reconcil.)
0.769
0.894
0.703
0.777
0.840
80.7
r2s_mix (rooms+reconcil.)
0.781
0.906
0.709
0.784
0.848
79.3
wf_cc5k (full seq)
0.747
0.762
0.742
0.374
0.341
117.1
wf_mix (full seq)
0.751
0.764
0.746
0.395
0.367
113.9
Table 4: Matched-data comparison, CubiCasa5K official test, fpeval @0.05. Room-centric rows are scored on the 290 (r2s_cc5k) and 287 (r2s_mix) plans on which the model produces a decodable output, wall-first rows on all 296 ; on the common 287 -plan roster wf_mix has wall F1 0.751 and edit cost 110.2 ( Table 11 ). “rooms” for wall-first is derived from faces; room-centric rooms are the Raster2Seq stage-1 output scored under fpeval against skv4, not under the official protocol (which gives 0.887 , Section 5 ). Per-system 95% bootstrap half-widths on wall F1 are 0.013 – 0.017 . Wall-first has higher wall recall; the room-centric system is better on wall F1, rooms, openings, watertightness, and edit cost.
arm
room-boundary walls ( n=3337 )
non-boundary walls ( n=835 )
r2s (rooms+reconcil.)
0.865
0.302
wall-first native
0.827
0.663
Table 5: Wall recall by room-boundary membership (150-plan subset of the CubiCasa5K official test, tol 0.05). The representational effect is a coverage limit rather than a topology-quality effect.
Figure 3: Wall classes. Room-bounding walls can be recovered from room polygons; non-room-bounding walls (dangling, stub, exterior) cannot.
system
wall F1
wall P
rooms
doubles/plan
rval
r2s_cc5k + reconcil.
0.769
0.894
0.777
0.00
0.938
r2s_cc5k NAIVE
0.708
0.758
0.778
7.90
0.983
wf_cc5k snap .06
0.747
0.762
0.374
∼ 0
0.02
wf_cc5k SNAP0
0.778
0.794
0.087
0.70
0.01
Table 6: Post-processing ablation, CubiCasa5K test, fpeval @0.05. Without the reconciliation step or snapping, each system is better on its own native element; the aggregate ordering is set by the asymmetry of derivation direction.
Table 7: Fusion steps, CubiCasa5K official test ( n=287 of 296), wall F1, with the donor stated per row. Each step requires no training and is deterministic. Opening-strip changes openings and edit cost only, not wall F1. The final system is the best overall (paired 95% interval of its gain over r2s_mix: [5.9,8.3] pp at 0.05 , [5.1,7.2] at 0.015 ).
domain
donor
base F1
F1/2
k
p
fused F1
Δ
CubiCasa5K
wf_mix seq, no gate
0.781
0.391
7.4
0.51
0.832
+5.1
CubiCasa5K
wf_mix seq, ink gate
0.781
0.391
4.8
0.67
0.846
+6.5
CubiCasa5K
walls-only seq, no gate
0.781
0.391
7.1
0.56
0.844
+6.3
CubiCasa5K
walls-only seq, ink gate
0.781
0.391
4.9
0.69
0.853
+7.1
CubiCasa5K
graph readout, no gate
0.781
0.391
4.8
0.75
0.856
+7.4
CubiCasa5K
graph readout, ink gate
0.781
0.391
4.5
0.74
0.850
+6.9
Table 8: Measured donor purity p against the threshold F1/2 of Eq. 4 , wall [email protected]. k is the mean number of ink-gated donor walls added per plan; p is the pooled fraction of those that match a ground-truth wall the base had missed. CubiCasa5K rows are on the 287 -plan roster; ResPlan is scored on the 733 zero-shot plans where the room-centric base decodes.
arm
deploy
zero-content
GT-content
no-conditioning (walls-only)
0.790 / 0.736
room-outline channel
0.762 / 0.708
0.762 / 0.710
0.764 / 0.709
fused-draft channel
0.768 / 0.720
0.792 / 0.742
0.795 / 0.745
symbolic coordinate prefix
0.756 / 0.696
0.783 / 0.717
0.781 / 0.715
Table 9: Input-side conditioning, CubiCasa5K official test, wall F1 (@.05 / @.015). The three columns for each arm are deployment / zero-content / ground-truth-content. No deployment arm beats the unconditioned wall-only model (0.790 / 0.736); the zero- and GT-content control rows reproduce that baseline within noise, and GT content adds at most 0.3 pt over zero-content. The paired noise scale on this test set is about ±1.3 pt ( Section 6.1 ), so the 0.3 pt control differences are indistinguishable from zero while the 2 – 3 pt deploy-arm deficits are not.
Figure 4: The fused-draft-channel conditioning arm. The rasterized draft enters the wall-first network as an extra input channel; the room-outline arm replaces the draft raster with rasterized room outlines, and the symbolic-prefix arm instead feeds the draft’s coordinates to the decoder as a prefix (not drawn). What any of them can carry is a second reading of the same pixels, and none has an effect.
track
score
note
zero-shot, fusion system
0.688
readout fallback on failures
zero-shot, wall-first seq-decode alone
0.640
no failures
zero-shot, wall-first readout alone
0.621
frozen thr.; 0.677 calibrated
zero-shot, room-centric alone
0.383
26.7% hard failures
finetuned, room-centric
0.928
rooms/open/wt 0.986/0.989/0.986
finetuned, wall-first seq-decode
0.968
—
Table 10: ResPlan-FP baseline tracks (wall F1 unless noted). The zero-shot fusion row keeps the strictly zero-shot protocol (frozen-threshold readout donor); fusion with the threshold-calibrated donor is an expected further gain and is not listed as a track.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
operation
r2s_mix
wf_mix
fusion
wall move
21.7
23.3
25.3
wall type change
8.5
2.5
10.3
wall create
7.5
6.0
4.0
wall delete
1.8
5.1
3.1
open-to-phantom convert
0.0
0.0
0.0
opening move
13.6
10.3
13.6
Appendix
Table 11: Mean edit operations per plan, CubiCasa5K official test ( n=287 ), default weights. Rows are operation counts; the total is the weighted cost of Eq. 1 , in which a move is weighted by its displacement, so the total is not a weighted sum of the rows. The wall-first total here ( 110.2 ) differs from Table 4 ( 113.9 , n=296 ) because of the roster; the room-centric total is unchanged because Table 4 already scores it on these 287 plans.
bed/bath
share
parameter
value
0/1 (studio)
0.5%
area band, 1 bed
30–80
1/1
15.2%
area band, 2 bed
50–130
2/2
28.4%
area band, 3 bed
75–200
3/2
36.0%
area band, 4 bed
110–350
4/3
16.8%
balcony
0.70
5/4
2.8%
service yard
0.60 ( ≥ 2 bed), 0.25 otherwise
Appendix
Table 12: Unit statistics used by the generator. Left: bedroom/bathroom joint distribution fitted to 18,203 parsed listings. Right: area bands ( m2 ) per bedroom count and prevalence of optional spaces.
Two structural insights have been overlooked in automated residential floor plan generation. First, design is inherently progressive. Architects begin with rough strokes and refine them over time, whereas existing methods typically require their conditioning representation to be fully specified before generation, a fundamental mismatch with how design actually works. Second, the 2D floor plan is not an optional intermediate but an irreplaceable spatial contract. Once room boundaries, doors, and windows are fixed, furnishing reduces from open-ended spatial reasoning to bounded constraint satisfaction. Bypassing this contract, as existing 3D systems do by delegating layout to language models, yields overlapping rooms and implausible proportions; directly calling general-purpose language models likewise produces geometrically invalid layouts. Guided by these insights, we present PlanCraft. SketchPlan supplies the missing training signal by replaying the architect's drawing process on 80K real floor plans, producing partial sketches at every completeness level. PlanCraft-Diff progressively sharpens an incomplete sketch into a geometrically precise, vectorizable floor plan through a coarse-to-fine strategy. With the spatial contract established, PlanCraft-Agent then furnishes the scene within well-defined room boundaries. Experiments show that PlanCraft achieves a 61.1% lower FID than the best existing 2D method and surpasses existing 3D systems by 15 points in expert-rated spatial rationality, with a sketch at only 25% completion already outperforming all fully specified baselines.
In this work, we propose HypergraphFormer, a novel and efficient approach to floor plan generation based on learning hypergraph representations with a large language model (LLM). The model is trained via supervised fine-tuning to generate a hypergraph-based textual representation that encodes spatial relationships and connectivity information within floor plans. We train and evaluate our approach on the RPLAN dataset, and further demonstrate its generalizability on a separate out-of-distribution dataset, which we release in this paper. Our method outperforms state-of-the-art techniques based on rasterized or vectorized representations across a diverse set of metrics. We also show improved data efficiency, particularly under distribution shift. The hypergraph formulation enables the generation of floor plans for arbitrary, irregular, user-specified boundaries by decoupling apartment footprints from their functional and geometric subdivisions. Furthermore, we show that the proposed methodology offers a high degree of editability, making it particularly well suited to design-oriented workflows supported by LLMs.
Estimating room layouts from multi-view imagery is a core task for indoor scene understanding. Existing methods are typically limited either by poor generalization to new datasets or restrictive geometric assumptions of the room shape or camera configuration. Most also estimate rooms independently, failing to exploit shared building structure such as dominant directions, ground plane or ceiling height. We propose PolyLayout, a multi-room layout estimation method that parameterizes room layouts as Manhattan 3D polygons and optimizes them jointly across multiple rooms. The optimization objective is predicted by a neural network on top of robust pre-trained visual features and trained end-to-end with supervision only on output room layouts. At the same time, camera projection and polygon updates remain explicit and model-based. This separation between learned scoring and geometry improves generalization to new datasets and camera parameters. During optimization, PolyLayout adaptively refines the polygon topology through iterative wall split and merge operations while jointly utilizing structural cues across rooms. We introduce two new multi-view multi-room layout benchmarks by providing layout annotations to existing datasets, and experiments show that PolyLayout outperforms prior approaches, both in terms of accuracy and robustness. Project page: https://ghanning.github.io/PolyLayout
Gustav Hanning, Shaohui Liu, Rémi Pautrat +3
Lund University · ETH Zurich · Microsoft Spatial AI Lab