Rethinking Floorplan Generation with Space Syntax: From Design Theory to Model Improvement

Zhuoyang Jiang; Dongqing Zhang; Qingxiang Liu; Yuxuan Liang

cs.LGFeb 26, 2026

Rethinking Floorplan Generation with Space Syntax: From Design Theory to Model Improvement

Authors: Zhuoyang Jiang, Dongqing Zhang, Qingxiang Liu, Yuxuan Liang

Organizations: College of Architecture and Urban Planning, Tongji University, Shanghai, China · Information Hub, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China

Abstract

Architectural design increasingly adopts generative AI (GenAI) for floorplan generation, an early-stage design task exploring how spaces should be organized for use. Yet fidelity objectives inherited from native visual and geometric generators do not automatically capture this design purpose. This organization remains largely implicit in generated representations, limiting direct evaluation and feedback. Drawing on space syntax, an established architectural design theory, the Space Syntax Integration Operator (SSIO) recovers convex-space access graphs across heterogeneous formats. This theory-grounded shared intermediate representation, which we call a design container, supports evaluation and model understanding of spatial design. Applying SSIO to four residential corpora spanning 15 countries reveals public spaces' integration levels are highest, with corpus-dependent distributions. These patterns underpin the RPLAN-SSI benchmark, an empirical design reference system assessing population-level functional-profile alignment (IPD) and within-plan public-space advantage (PIA). Comparison with this residential reference reveals a design gap: public spaces in generated floorplans have a smaller integration advantage over other spaces. To address this gap, Space Syntax-Guided Post-Training (SSPT) provides a plug-and-play design-learning interface: SSIO-derived RL feedback improves IPD and PIA across five heterogeneous diffusion backbones while preserving native mechanisms. We further discuss how this approach may inform GenAI applications in other design domains.

Figures & tables

Figure 1: Towards Design-directed Representation and Measurement in Floorplan Generation.

Figure 2: Space Syntax Integration Operator (SSIO). Recovering shared theory-defined IR.

Figure 3: Cross-Corpus Grounding: Public-first.

Figure 4: Preserved Public-first Ordering, Weaker Functional Hierarchy.

Figure 5: SSPT-PPO trajectory-space post-training. Guarded SSIO rewards are clipped and standardized over rollout pools; the resulting trajectory advantages drive PPO updates on backbone-native lower-noise transitions.

Method	IPD $\downarrow$	median PIA $\uparrow$	Acc. (%)
CE2EPlan	0.1914	0.0726	66.5
DiffPlanner	0.1803	0.1113	59.1
House-GAN++	0.1612	0.1583	40.7
House-GAN	0.1377	0.2606	36.1
Graph2Plan	0.1078	0.2886	71.6
FloorplanGAN	0.0943	0.2296	91.5

Table 1: RPLAN-SSI evaluation.

Figure 6: Design evaluation and post-training. (a) Model families and Exact-8 acceptance (bubble area). (b) Gap closure to zero IPD and RPLAN median PIA. (c) AreaGuard trade-off: independent Base evaluation, Base-calibrated limits (dashes), and standard deviations (error bars).

Variant	FID $\downarrow$	Compatibility $\downarrow$
Base HD	$7.798\pm 0.140$	$2.537\pm 0.026$
+SSPT-PPO	$9.087\pm 0.098$	$2.572\pm 0.049$
Upper limit	$9.500$	$2.800$

Table 2: Native-quality trade-off of the selected HD checkpoint. Five-repeat mean

\pm

std.

Figure 7: HD+SSPT-PPO reorganizes functional configuration under fixed programs. Top: PIA changes for 4,588 matched pairs. Bottom: ten examples; heatmaps encode relative convex-space integration and links indicate room-level access.

Reverse-time support	IPD $\downarrow$	median PIA $\uparrow$
Lower-noise segment	0.0945	0.3349
Uniform full horizon	0.1026	0.2996
Higher-noise segment	0.1126	0.2781

Table 3: HouseDiffusion reverse-time ablation.

Variant	IPD $\downarrow$	median PIA $\uparrow$	FID $\downarrow$	Compatibility $\downarrow$
Anchor	0.0830	0.3912	$19.55\pm 0.18$	$2.583\pm 0.041$
+AreaGuard-2	0.0827	0.3921	$8.91\pm 0.12$	$2.118\pm 0.040$
+AreaGuard-4	0.0858	0.3817	8.59 $\pm$ 0.11	2.062 $\pm$ 0.063

Table 4: AreaGuard improves the quality–design trade-off. Iteration-10 development.

	IPD $\downarrow$		median PIA $\uparrow$
Backbone	Base	+SSPT-PPO	Base	+SSPT-PPO
HD	0.1111	0.0811	0.2800	0.3974
CE2EPlan	0.1914	0.1352	0.0726	0.2441
DiffPlanner	0.1803	0.1555	0.1113	0.1721
ChatHD	0.0866	0.0838	0.3295	0.3336
GSDiff-SSIO	0.0386	0.0360	0.5766	0.5934

Table 5: SSPT-PPO across five diffusion backbones.

Appendix figures & tables46 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
Design recovery and measurement
$c=(B,\mathcal{M}_{c},G_{c})$	boundary, requested room multiset, optional relation graph
$\mathcal{A}_{m}$	deterministic native-output normalization internal to SSIO
$\mathcal{C}_{\phi}$	frozen access-completion map when native access is absent
$z_{p,m}$	canonical measurement record for plan $p$ and method $m$
$G_{p,m}=(V_{p,m},E_{p,m})$	room-labeled rectangular convex-atom access graph

Appendix

Table 6: Notation used in the paper.

Canonical category	Source labels	Benchmark role
Living	Living	Public
Dining	Dining	Public
Bedroom	Master, Child, Second, Guest	Private
Study	Study	Private
Kitchen	Kitchen	Service
Bathroom	Bathroom	Service

Appendix

Table 7: Semantic crosswalk for the RPLAN benchmark.

Quantity	Aggregation or contrast	Role
Functional profile $\widetilde{\mathbf{r}}$	Within-plan relative domain scores, then cohort medians	Population organization; IPD
Public-domain margin $\Delta_{\mathrm{pub}}$	Public minus strongest competing relative domain score	Individual domain hierarchy
PIA	Living category mean minus strongest eligible category mean; category-normalized	Individual RPLAN anchor
$R_{\mathrm{anchor}}$	Largest Living room score minus largest non-Living room score; unnormalized	Post-training feedback

Appendix

Table 8: Distinct summaries and learning signals derived from SSIO scores.

Role	Size	Use
RPLAN real-corpus prior	80,788	cross-corpus functional regularity
LIFULL Living/Dining cohort	19,677 / 19,541	relative-domain-score / margin support
MSD Swiss corpus	13,472	cross-corpus functional regularity
SFC-A68 multi-country corpus	270	cross-corpus functional regularity
RPLAN exact-eight reference	20,348	frozen four-domain benchmark profile
SSPT-PPO rollout conditions	$\leq 7$ -room requests	post-training only

Appendix

Table 9: Data roles in analysis and evaluation. Development, validation, calibration, and the endpoint pool below refer to the primary HD study; Table 11 gives the other backbones’ training scales.

Figure 8: SSPT-Iter: design feedback through data selection. Generated candidates and a fixed real replay pool are ranked by SSIO scores. The top- $K$ selection supplies available targets for ordinary diffusion retraining before the next round.

Backbone	Native variable	Active coordinates	Trainable component
HouseDiffusion	$2\times 100$ coordinate tensor	all 200 coordinates	denoiser
CE2EPlan	$6\times L_{m}$ room tensor $[\mathrm{pres},\mathrm{type},x,y,w,h]$	unpadded $x,y,w,h$ coordinates	denoiser
DiffPlanner	$4\times L_{m}$ room-box tensor $xyxy$	unpadded room boxes	PartitioningDiff
ChatHD	$1\times 64\times 64$ semantic raster	interior raster pixels	suffix U-Net
GSDiff–SSIO	$53\times 10$ corner/semantic tensor	joint reverse tensor	main diffusion model

Appendix

Table 10: Native policy bindings used by the five reported diffusion backbones.

Backbone	Iter. $\times$ rollouts/iter.	PPO epochs	LR	$M_{m}$	Atom area (px 2 )
HouseDiffusion	$20\times 10{,}000$	2	$5\times 10^{-6}$	10	50
CE2EPlan	$20\times 128$	2	$5\times 10^{-6}$	10	16
DiffPlanner	$60\times 512$	4	$10^{-6}$	10	16
ChatHD	$5\times 512$	2	$5\times 10^{-6}$	10	50
GSDiff–SSIO	$10\times 512$	2	$5\times 10^{-6}$	20	16

Appendix

Table 11: Numerical settings of the backbone training bindings.

	FID			Compatibility
Base batch	Mean	$s$	$U_{b,j}$	Mean	$s$	$U_{b,j}$
A: loader-default	8.5157	0.1670	8.9342	2.5240	0.0650	2.6869
B: loader-default	8.6177	0.3220	9.4243	2.5066	0.0207	2.5586
V: native validation	7.4739	0.0996	7.7234	2.6020	0.0719	2.7822
Adopted upper limit	9.5			2.8

Appendix

Table 12: Base-only calibration of the native-quality range. Each entry uses five 1,000-condition repeats. Standard deviations

s

use denominator four. The upper bounds jointly cover six future Base means under Eq. ( C.6 ); bold entries determine the common ceilings.

Centrality	Cohort	RPLAN/LIFULL $N$	RPLAN med.	LIFULL med.	Ratio
Legacy HH	all eligible	80,788 / 19,541	0.7162	0.3052	42.61%
Legacy HH	connected only	73,997 / 6,625	0.7062	0.3013	42.66%
Bounded HH, $n>2$	all eligible	80,710 / 18,554	0.2602	0.1210	46.48%
Bounded HH, $n>2$	connected only	73,997 / 6,590	0.2583	0.1268	49.08%
Global harmonic	all eligible	80,788 / 19,541	0.2791	0.1086	38.90%
Global harmonic	connected only	73,997 / 6,625	0.2759	0.1048	37.99%

Appendix

Table 13: Robustness of

\Delta_{\mathrm{pub}}

in RPLAN and LIFULL across centrality and connectivity choices.

Centrality	Public-margin deficit	IPD Base $\rightarrow$ +	median PIA Base $\rightarrow$ +
Legacy HH	46.27%	0.1111 $\rightarrow$ 0.0811	0.2800 $\rightarrow$ 0.3974
Bounded HH	38.68%	0.0368 $\rightarrow$ 0.0230	0.1060 $\rightarrow$ 0.1573
Bounded HH, $n>2$	38.54%	0.0376 $\rightarrow$ 0.0223	0.1060 $\rightarrow$ 0.1576
Global harmonic	42.36%	0.0503 $\rightarrow$ 0.0178	0.1186 $\rightarrow$ 0.2022

Appendix

Table 14: Centrality sensitivity of the task-matched functional gap and SSPT-PPO gains.

	Configurational metrics		Protocol outcome
Method	Block IPD $\downarrow$	Block median PIA $\uparrow$	Accepted
CE2EPlan	0.1916 $\pm$ 0.0020	0.0745 $\pm$ 0.0114	3,324/5,000
DiffPlanner	0.1806 $\pm$ 0.0028	0.1119 $\pm$ 0.0080	2,956/5,000
House-GAN++	0.1611 $\pm$ 0.0034	0.1599 $\pm$ 0.0100	2,036/5,000
House-GAN	0.1371 $\pm$ 0.0047	0.2616 $\pm$ 0.0080	1,803/5,000
Graph2Plan	0.1074 $\pm$ 0.0018	0.2899 $\pm$ 0.0095	3,582/5,000

Appendix

Table 15: Mean and population standard deviation across five evaluation blocks for the benchmark and primary post-training result. The main table pools all five blocks.

Statistic	Base	+SSPT-PPO
median PIA [25th, 75th percentiles]	0.2809 [0.1306, 0.3937]	0.3968 [0.2870, 0.4916]
$\Pr[\mathrm{PIA}>10^{-6}]$	87.90%	98.10%
IPD on joint subset	0.1112	0.0812

Appendix

Table 16: Condition-matched sensitivity on 4,588 jointly accepted outputs.

Method	Conn./accepted	Conn./attempted	IPD	median PIA
Base	4,223/4,723 (89.41%)	4,223/5,000 (84.46%)	0.1124	0.2742
+SSPT-PPO	4,311/4,825 (89.35%)	4,311/5,000 (86.22%)	0.0833	0.3883

Appendix

Table 17: Single-component sensitivity on the five-block benchmark.

Source	Public	Private	Service	Outdoor
RPLAN reference	1.5548	0.8772	0.7860	0.7445
HouseDiffusion Base	1.3465	0.9565	0.9210	0.7664
HouseDiffusion+SSPT-PPO	1.4044	0.9068	0.8686	0.8064

Appendix

Table 18: Median relative-integration profile coordinates.

	Configurational metrics		Protocol outcome
Reward	IPD $\downarrow$	median PIA $\uparrow$	Accepted
Anchor	0.0830	0.3912	4753/5000
Category alignment	0.0828	0.3903	4782/5000
Living/MAD	0.0927	0.3554	4747/5000
Anchor+category alignment	0.0807	0.4027	4764/5000

Appendix

Table 19: Reward identity on the development pool at iteration 10.

Figure 9: AreaGuard strength at a fixed training duration. Anchor and iteration 10 are held fixed. Weight 2 improves all four plotted metrics relative to no guard; weight 4 further favors native quality at some design-score cost. Design metrics use 5,000 development conditions; native metrics use five repeats of 1,000 conditions. Error bars show repeat standard deviations; dashed lines mark the calibrated Base-reference ceilings.

Figure 10: What changes near the Base-referenced FID ceiling? Replacing Base outputs with same-condition Anchor outputs increases geometric extent and eventually FID. Curves show means across three substitution orders; bands span their minima and maxima. Dashed lines mark FID 9.5 and unit geometric ratios; dotted FID lines show alternative ceilings. Ratios are averaged per-image ratios to the matched real rendering.

Guard	Iter.	IPD $\downarrow$	median PIA $\uparrow$	Acc.	FID $\downarrow$	Comp. $\downarrow$	Status
2	10	0.0842	0.3934	0.9665	8.594	2.210	Eligible
2	15	0.0812	0.4031	0.9610	10.061	2.430	Excluded
2	19	0.0778	0.4067	0.9560	12.330	2.640	Excluded
2	22	0.0811	0.4050	0.9510	13.557	2.787	Excluded
2	25	0.0822	0.4040	0.9445	15.084	2.967	Excluded
2	30	0.0835	0.3913	0.9210	19.145	3.343	Excluded

Appendix

Table 20: Validation separates eligibility from design-based selection. Native metrics are five-repeat means; acceptance is a fraction. All candidates satisfy support and acceptance requirements. Status applies the Base-calibrated FID/Compatibility ceilings of 9.5/2.8; the selected row is bold.

Figure 11: Training duration and the quality–design trade-off. Lines connect the nine measured validation checkpoints in Table 20 ; native-metric error bars show repeat standard deviations. Shading marks iterations 15–20, where weight 4 yields smaller additional design gains at increasing native cost. The star identifies the selected checkpoint. Dashed lines mark the Base-calibrated ceilings.

Figure 12: Checkpoint selection under different quality requirements. (a) Validation FID and Compatibility; shading marks the admissible region and error bars show repeat standard deviations. (b) Selection with the candidate set and Eq. ( C.8 ) fixed; labels denote guard weight/iteration. The star marks the Base-calibrated ceilings, 9.5/2.8, selecting weight 4 at iteration 20. Stricter ceilings can select iteration 15.

Backbone	Variant	IPD $\downarrow$	median PIA $\uparrow$	Accepted
HouseDiffusion	Base	0.1111	0.2800	4,723/5,000
	+SSPT-PPO	0.0811	0.3974	4,825/5,000
CE2EPlan	Base	0.1914	0.0726	3,324/5,000
	+SSPT-PPO	0.1352	0.2441	2,956/5,000
DiffPlanner	Base	0.1803	0.1113	2,956/5,000
	+SSPT-PPO	0.1555	0.1721	3,227/5,000

Appendix

Table 21: Full paired portability results, including Exact-8 acceptance.

Criterion	Required	Observed	Met
Extra pairs introduced by opening realization	$=0$	0	Yes
Plans with opening-realization issues	$\leq 0.020$	0.028	No
Access-pair precision	$\geq 0.900$	0.8121	No
Access-pair recall	$\geq 0.900$	0.9792	Yes
Access-pair F1	$\geq 0.900$	0.8879	No
Unambiguous-candidate recall	$\geq 0.995$	1.000	Yes

Appendix

Table 22: Access-completion fidelity on 500 held-out plans. Thresholds were fixed before evaluation. Pair recovery, opening realization, and preservation of integration are tested separately.

	IPD $\downarrow$		median PIA $\uparrow$
Predictor features / decoding	Base	+SSPT-PPO	Base	+SSPT-PPO
Full / connected	0.0362	0.0351	0.5882	0.5984
Full / raw	0.0275	0.0302	0.6144	0.6299
Geometry/context / connected	0.1102	0.1087	0.2781	0.2877
Type / connected	0.0407	0.0414	0.6093	0.6165

Appendix

Table 23: Access-source attribution on 538 common conditions from a separate 1,000-condition experiment. Feature restrictions apply to the frozen pair predictor; every route retains geometric candidate and opening construction.

Figure 13: Research emphasis across all 55 reviewed studies. Publication setting and contribution type are classified independently of the investigated objective. A study may contribute to multiple types in panel B. Study-level assignments and references appear in Section F.6.2 ; Figure 14 explains representative mechanisms.

Figure 14: Research setting, computational intervention, and validation. Representative cases explain the 55-study distribution in Figure 13 ; the small-residence analysis adds architectural context outside that coded sample. References and validation details appear in Section F.6.2 and Table 25 .

Design object	Synthesis-oriented question	Purpose-oriented question
Graphic layout	Are elements and placements plausible?	Does their organization establish a readable hierarchy?
Mechanical part	Is the geometry valid and similar to examples?	Does the structure perform under its intended loads?
Garment	Can a coherent pattern be generated?	Do panel and sewing relations support assembly and wearing?
Floorplan	Are rooms and boundaries geometrically coherent?	Does their access organization support the intended use of spaces?

Appendix

Table 24: Synthesis quality and design purpose ask different questions about the same artifact.

Figure 15: Representation and criterion use are distinct. Named studies connect represented organization to criterion use. Figure 17 distinguishes theory-derived measures from other criteria; Section F.6.2 gives the complete inventory.

Figure 16: Knowledge provenance and computational destination across all 55 studies. Cells count documented source-to-use links, including learned implementations; both dimensions may overlap. Marked cells include abstract-level or contextual evidence. Our framework is excluded from these counts and appears separately in Figure 17 .

Figure 17: Representative knowledge-to-computation connections. These cases explain the categories in Figure 16 . Our pathway carries space-syntax objects and measurements through empirical grounding into heterogeneous-generator post-training. Study references are given in the accompanying analysis and inventory.

Figure 18: From design signals to model learning. Fifty studies are classified by (A) analysis interface and signal basis and (B) the component affected. Organization/performance includes authored criteria and physical objectives as well as theory-derived measures; Figure 17 separates these foundations. The star denotes our SSIO/RPLAN-SSI/SSPT pathway. The key names each study; Section F.6.2 provides references.

Figure 19: Post-training scope and feedback interface. Seven prior methods and our pathway are classified by their feedback basis and demonstrated generator adaptation. IR denotes intermediate representation. SSIO and SSPT connect theory-defined analytical objects to native policy updates across heterogeneous outputs.

Figure 20: A design criterion can update different components. The examples separate native training, representation or evaluator learning, prompt-policy learning, and generator post-training. SSPT connects one architectural measurement to heterogeneous diffusion policies.

Study / community	Reported scale or protocol	What the evidence tests
DRL-Urban / interdisciplinary ( Zheng et al., 2023 )	One synthetic scenario and two real communities; repeated policy runs; eight professional-planner comparisons; 100 design-trained evaluators.	Use-related planning objectives and professional comparisons across the evaluated settings.
Museum-CGAN / interdisciplinary ( Tang et al., 2024a )	100 images and 264 valid questionnaire responses; no conventional train/test split reported.	Perceived design outputs; questionnaire scale does not establish held-out generator generalization.
Small-residence analysis / specialist journal ( Man et al., 2026 )	14,373 extracted small-layout samples; evaluation demonstrated on a 30 m 2 residential unit.	Dataset analysis and a design-assessment case, rather than large-scale model adaptation.
Rule-Exhibit / graphics ( Xu et al., 2025 )	Rule-generated supervision and quantitative comparisons; dataset size and split not established in the inspected report.	A rule-to-learning mechanism; the available report leaves independent testing breadth unspecified.
Data Formulator 2 / HCI ( Wang et al., 2025 )	Eight participants in a user study.	Iterative analytical workflow and interaction.
LaySPA / AI ( Li et al., 2026b )	3,000 training layouts per dataset; 1,000 and 905 test layouts.	Generator post-training and held-out layout criteria.

Appendix

Table 25: Validation protocols answer different questions. Source-data volume, independent design cases, and human assessments are reported separately.

Study / source	Representation and criterion	Where the criterion acts
LayoutTransformer ( Gupta et al., 2021 ) G01; AI; R1, P0 F:S; K:E	Element boxes and labels. Layout distribution and completion.	Autoregressive likelihood fitting.
LayoutDM ( Inoue et al., 2023 ) G02; AI; R2, P2 F:S; K:R	Elements with operational pairwise constraints. Size / position relation satisfaction.	Denoising; relation energies guide sampling.
LayoutFormer++ ( Jiang et al., 2023 ) G03; AI; R2, P2 F:S; K:R	Serialized element and relation constraints. Constraint satisfaction.	Likelihood fitting; restricted decoding.
LayoutNUWA ( Tang et al., 2024b ) G04; AI; R1, P1 F:S; K:A	Layout attributes in editable code. Alignment and overlap, evaluated.	Reference-code instruction tuning.
PosterLLaVa ( Yang et al., 2024b ) G05; PP; R1, P1 F:B; K:A+R	Poster elements as JSON boxes. Saliency-aware placement, evaluated.	Visual instruction tuning to reference layouts.
LaySPA ( Li et al., 2026b ) G06; AI; R2, P3 F:B; K:A+R	Explicit element environment and relations. Spacing, alignment and text-underlay pairing.	GRPO updates the LLM layout generator.

Appendix

Table 26: Layout generation and spatial control in graphic design.

Study / source	Representation and criterion	Where the criterion acts
ChartGPT ( Tian et al., 2025 ) G08; HG; R2, P2 F:B; K:R	Data transformations and chart specifications. Valid data fields and chart encodings.	Specification SFT; inference validity filtering.
Data Formulator 2 ( Wang et al., 2025 ) G09; HG; R2, P2 F:D; K:R	Data transformations and encoding specifications. Iterative analytical task completion.	User and LLM revise data / code; no new weights.
AutomaTikZ ( Belouadi et al., 2024 ) G10; AI; R2, P2 F:S; K:R	Executable TikZ programs. Compilable scientific vector graphics.	Code SFT; compiler-guided retry.
DiagramAgent ( Wei et al., 2025 ) G11; AI; R2, P2 F:S; K:A+R	Diagram plans and executable diagram code. Completeness and compilability.	Code-agent SFT; inference checking.
AutoFigure ( Zhu et al., 2026 ) G12; AI; R2, P2 F:B; K:A+R	Entity-relation blueprint and rendered diagram. Scientific fidelity and diagram clarity.	Inference critique and refinement.
UIClip ( Wu et al., 2024a ) G13; HG; R0, P3 F:D; K:P+A+J	Screenshots and text assessments. Contrast, repetition, alignment, proximity.	CLIP evaluator learns perturbations / preferences.

Appendix

Table 27: Intermediate specifications for visualization, diagrams, and interface evaluation.

Study / source	Representation and criterion	Where the criterion acts
UICoder ( Wu et al., 2024b ) G14; AI; R2, P3 F:S; K:R	SwiftUI code and rendered interfaces. Compilation and prompt relevance.	Filtered SFT and DPO update code generator.
IDEA-Bench ( Liang et al., 2025 ) E01; AI; Context F:B; K:R+J	Images and task-specific settings. Professional task completion and quality.	Benchmark; no generator update. Not in coordinate map.
Q-Judger ( Li et al., 2026a ) E02; PP; R0, P3 F:B; K:R+J	Images and explicit rubric text. 56 professional image-evaluation facets.	Supervised Q-Judger evaluator training.
ArtiMuse ( Cao et al., 2026 ) E03; AI; R0, P3 F:D; K:J	Images with aesthetic analyses. Eight aesthetic attributes and holistic judgment.	Expert-supervised aesthetic evaluator.
GDB-Reward ( Deganutti et al., 2026 ) E04; PP; R1, P3 F:B; K:A+R+J	Layout metadata, prompts and images. Text / color / layout reference fidelity.	GRPO updates prompt policy; FLUX remains frozen.
PosterReward ( Lai et al., 2026 ) E05; AI; R0, P3 F:B; K:A+J	Poster images and learned textual judgments. Typography, composition and preferences.	Reward learning; downstream generator RL adaptation.

Appendix

Table 28: Interface generation and professional design evaluation.

Study / source	Representation and criterion	Where the criterion acts
DeepCAD ( Wu et al., 2021 ) P01; AI; R2, P1 F:S; K:R	Sketch / extrusion command sequences. CAD execution validity.	Command and parameter fitting.
BrepGen ( Xu et al., 2024 ) P02; HG; R2, P2 F:S; K:R	Face-edge-vertex B-rep structure. Geometrically valid assembly.	Diffusion fitting; geometry repaired at inference.
Text2CAD ( Khan et al., 2024 ) P03; AI; R2, P1 F:B; K:R+J	Text-conditioned CAD programs. Valid execution and prompt correspondence.	Conditional CAD sequence fitting.
TopoDiff ( Mazé & Ahmed, 2023 ) P04; AI; R1, P2 F:B; K:R+M	Material occupancy and load fields. Compliance and absence of floating material.	Surrogate gradients guide diffusion sampling.
MO-PaDGAN ( Chen & Ahmed, 2021 ) P05; IX; R1, P3 F:B; K:A+M	Parametric engineering shapes. Performance and diversity.	Additional objectives train GAN generator.
DressCode ( He et al., 2024 ) P06; HG; R2, P1 F:S; K:R+J	Panels and sewing connections. Draping and prompt correspondence, evaluated.	Autoregressive pattern fitting.

Appendix

Table 29: CAD, engineering performance, and garment assembly.

Study / source	Representation and criterion	Where the criterion acts
AIpparel ( Nakayama et al., 2025 ) P08; AI; R2, P0 F:S; K:E	Multimodal sewing-pattern tokens. Paired panel and stitch reconstruction.	Supervised multimodal generation / editing.
TextileDiffusion ( Karagoz et al., 2023 ) P09; DS; R0, P0 F:S; K:E	Textile-pattern images. Appearance and text similarity.	Image-domain diffusion adaptation.
Knit-FLUX ( Liu et al., 2025c ) P10; IX; R0, P1 F:S; K:A+J	Rendered knit images. Knit-unit appearance and yarn coverage, evaluated.	Image-domain LoRA adaptation.
FunctionalTextile ( Tang et al., 2026 ) P11; DS; R2, P3 F:B; K:R+M	Reported yarn / interlacing fields. Reported porosity, stiffness and conductivity.	Reported functional-surrogate loss trains generator.
KumisPackaging ( Zhao et al., 2026 ) P12; IX; R1, P2 F:D; K:R+J	Cultural elements and shape-grammar variants. Cultural communication and packaging preference.	Human prompting and candidate selection.
CraftBeerPackaging ( Chen & Jang, 2026 ) P13; DS; Abstract only F:D; K:U	Packaging images from existing tools. Visual and brand-expression assessment.	Tool comparison; abstract-level evidence. Not in coordinate map.

Appendix

Table 30: Garments, textiles, packaging, and engineering-design metrics.

Study / source	Representation and criterion	Where the criterion acts
UrbanGAN ( Fedorova, 2021 ) S01; DS; R1, P1 F:S; K:A+R	Urban-block rasters and parcel context. Building spacing and urban morphology, evaluated.	Conditional adversarial fitting.
DRL-Urban ( Zheng et al., 2023 ) S02; IX; R2, P3 F:D; K:P+R	Parcel-road-intersection graph. Service access, ecology and road efficiency.	PPO trains a planning policy from scratch.
PlantoGraphy ( Huang et al., 2024 ) S03; HG; R2, P2 F:B; K:R+J	Scene graph and plant-layout constraints. User-specified landscape composition.	Interactive graph planning; rendering LoRA.
GenFusion-Urban ( Gao et al., 2025 ) S04; DS; Abstract only F:B; K:R+M	Reported urban layout / performance pipeline. Urban building performance.	Reported diffusion + surrogate + NSGA-II. Not in coordinate map.
TaleCrafter ( Gong et al., 2023 ) S05; HG; R1, P2 F:S; K:R	Character attributes, layouts and story sequence. Character consistency and story visualization.	Adapted rendering plus inference control.

Appendix

Table 31: Urban, landscape, and story visualization.

Study / source	Representation and criterion	Where the criterion acts
CineVerse ( Phung et al., 2026 ) S06; AI; R2, P2 F:B; K:P+R	Structured cinematic shot plans. Cinematic composition and continuity.	Plan-guided inference; keyframe fitting.
STAGE ( Zhang et al., 2026b ) S07; AI; R2, P3 F:B; K:A+J	Intra-shot and inter-shot frame pairs. Cinematic transition preference.	DPO updates frame-pair predictor; video model frozen.
ShotPlan ( Guo et al., 2026 ) S08; PP; R2, P1 F:B; K:R+J	Explicit shot / cut planning tokens. Shot controllability and consistency, evaluated.	Conditional flow matching; no separate planning loss.
Cont-PCGRL ( Earle et al., 2021 ) S09; AI; R2, P3 F:D; K:A+R	Grid regions, paths and edit actions. Target difficulty / path and region objectives.	From-scratch goal-conditioned PPO.
DiffusionCraft ( Dai et al., 2024 ) S10; AI; R1, P1 F:B; K:R	Level tiles and local spatial patterns. A-star playability, evaluated.	Diffusion fitting from one level.

Appendix

Table 32: Cinematic organization and playable levels.

Study / source	Representation and criterion	Where the criterion acts
PCGRL-Jax ( Earle et al., 2024 ) S11; AI; R2, P3 F:D; K:A+R	Grid edits, goals and fixed elements. Path / region control and generalization.	From-scratch PPO with scalable environments.
IPCGRL ( Baek et al., 2025 ) S12; AI; R2, P3 F:D; K:A+R	Language-derived tasks and level edits. Instruction-specified paths and regions.	Task encoding plus reward-trained PPO.
Museum-CGAN ( Tang et al., 2024a ) S13; IX; R1, P1 F:B; K:J	Museum zones and rendered plans. Spatial / visual qualities, evaluated.	Adversarial and perceptual image fitting.
Rule-Exhibit ( Xu et al., 2025 ) S14; HG; R2, P3 F:B; K:A+R	Hall connectivity and exhibit placement. Connectivity, saliency and exhibit importance.	Design rules generate supervised training labels.
Museum-LoRA-Compare ( Li et al., 2026c ) S15; DS; Abstract only F:B; K:T+J	Generated museum-plan images. Spatial connectivity, evaluated.	LoRA / GAN comparison; abstract-level evidence. Not in coordinate map.

Appendix

Table 33: Controllable levels and exhibition layouts.

Study / source	Representation and criterion	Where the criterion acts
House-GAN++ ( Nauata et al., 2021 ) F01; AI; R2, P2 F:S; K:R	Room graph and room masks. Graph compatibility and geometric quality.	GAN training; schedule meta-optimization.
HouseDiffusion ( Shabani et al., 2023 ) F02; AI; R2, P1 F:S; K:R	Graph-conditioned vector polygons. Geometry and compatibility, evaluated.	Noise and binary-coordinate regression.
ChatHouseDiffusion ( Qin et al., 2024 ) F03; PP; R2, P2 F:S; K:R	Language-derived relations and raster plan. User conditions and controlled editing.	Conditional denoising and inference editing.
GSDiff ( Hu et al., 2025a ) F04; AI; R2, P3 F:S; K:R	Boundary graph and vector geometry. Alignment and edge validity.	Native geometric losses train generator.
FloorPlan-LLaMa ( Yin et al., 2025 ) F05; AI; R0, P3 F:B; K:J	Image tokens; image-based preference evaluator. Expert functionality, flow and overall judgment.	PPO updates autoregressive image-token generator.
Floorplan RLVR ( Lara et al., 2026 ) F06; AI; R2, P3 F:S; K:R	JSON polygons and reconstructed relations. Connectivity-graph and area requirements.	GRPO updates autoregressive generator.

Appendix

Table 34: Floorplan synthesis and design-directed learning.

Figure 21: A proposed workflow for design theory translation. Informed by the floorplan study, the overview connects formalization, implementation, and cross-domain extension. Stage 1 defines typed analytical objects and quantitative criteria grounded in domain meaning. Stage 2 recovers the design container from artifacts, grounds measurements in empirical references, and connects shared feedback to native model updates.

Design category	Available foundation	Construction focus
Existing quantitative or formal analytical accounts
Floorplans / interiors	Space syntax	Configuration recovery and grounding
Urban / landscape	Space syntax; cooling models	Purpose-specific network or climate inputs
UI / interaction	Keystroke-Level Model	Executable methods and calibrated timings
Products / CAD	Axiomatic Design	Functional dependencies and success evidence
Visualization / maps	Expressiveness; projection theory	Task-matched encodings and distortion

Appendix

Table 35: Construction entry points across fourteen design categories. Groups describe the selected knowledge sources and purposes; the paragraphs below provide their evidence and operator proposals.

Figure 22: Worked theory-translation examples. (a) Task completion qualifies KLM time comparisons; M/H/P/K denote mental preparation, homing, pointing, and keystrokes. (b) Two editors start from the same development poster. Colors identify intended groups; image contents appear as labels in the recovered container. (c) A process recovered from the libSBGN glycolysis example; dashed arcs mark deleted production relations.

Selection criterion	Group disagreement $\downarrow$	Grid error $\downarrow$	Named panel $\uparrow$
No editing	0.350	0.463	58.3%
Geometry only	0.350	0.093	58.3%
Grouping feedback	0.083	0.278	79.2%

Appendix

Table 36: Poster repair under a shared editing budget. Means over four authored evaluation briefs. Named-panel accuracy is evaluated separately from the grouping objective.

Floor Plan Generation

See all 14 papers

PlanCraft: Sketch, Refine, and Furnish for Architect-Inspired Progressive 3D Residential Scene Generation
Jul 26, 2026Pengyu Zeng, Yuqin Dai, Jun Yin +73D Scene Generation 3D Scene Layout Generation
Tsinghua University
What a Comfortable World: Ergonomic Principles Guided Apartment Layout Generation
Apr 9, 2026Piotr Nieciecki, Aleksander Plocharski, Przemyslaw MusialskiFloor Plan Generation Generative Modeling
Warsaw University of Technology, Poland · Akces NCBR, Poland · New Jersey Institute of Technology, United States of America
HypergraphFormer: Learning Hypergraphs from LLMs for Editable Floor Plan Generation
May 18, 2026Nikita Klimenko, Hesam Salehipour, Parham Eftekhar +2Floor Plan Generation Indoor Scene Generation
Autodesk Research · York University · UC Berkeley

Jul 26, 2026J/K move · Enter open · S save

PlanCraft: Sketch, Refine, and Furnish for Architect-Inspired Progressive 3D Residential Scene Generation

Pengyu Zeng, Yuqin Dai, Jun Yin, Ziyang Han+6

Tsinghua University

3D Scene Generation 3D Scene Layout Generation Floor Plan Generation Indoor Scene Generation

Open paper AskPDF arXiv

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.

Abstract

Floor Plan Generation

Explore similar work

PlanCraft: Sketch, Refine, and Furnish for Architect-Inspired Progressive 3D Residential Scene Generation

What a Comfortable World: Ergonomic Principles Guided Apartment Layout Generation

HypergraphFormer: Learning Hypergraphs from LLMs for Editable Floor Plan Generation