acceptodds
Under review as a conference paper at ICLR 2027

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

18 views

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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Generative AI for Design, Design Theory Translation, Floorplan Generation, Space Syntax, Diffusion Model Post-Training, Reinforcement Learning for Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TnoWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how space syntax can be used to improve the spatial organization of floorplans generated by generative models. The authors propose SSIO, an analysis pipeline that converts heterogeneous floorplan representations into convex-space access graphs and computes spatial integration scores. Based on real residential floorplan datasets, they construct the RPLAN-SSI benchmark, using IPD and PIA to compare generated floorplans with real residential plans in terms of functional integration profiles and the integration advantage of public spaces. They further use integration-based rewards for PPO post-training and evaluate the approach on five diffusion-based floorplan generation models.

Critical strengths

The paper abstracts the theoretical account of spatial accessibility in space syntax, a theory from architectural design, into a unified convex-space access-graph representation and further uses this representation for diffusion-model post-training, offering cross-disciplinary insight. Building on this architectural design principle, the authors propose a set of quantitative metrics that transform the qualitative description of spatial organization in architectural theory into quantifiable statistics and training feedback, with potential cross-disciplinary applications.

Critical weaknesses

I have reservations about interpreting the reported metric improvements as evidence of design improvement. The core issue is that the paper uses the spatial organization property defined by the SSIO access graph and its integration scores both to construct the training reward and to define the main evaluation metrics, IPD and PIA. Although the training reward, IPD, and PIA are not identical formulas, they are all based on integration scores derived from the SSIO access graph. The training reward optimizes the integration difference between living rooms and nonliving rooms, PIA measures the difference between living rooms and other room categories, and IPD compares functional domain distributions aggregated from these scores. Therefore, the current experiments primarily demonstrate improvement in a proxy metric based on integration, rather than independently demonstrating improved architectural design quality or design capability. Comparisons on the same conditions and experiments using different centrality definitions strengthen the internal robustness of the metric improvements, but they do not resolve the issue of external validity for this proxy metric.

The paper’s construction of access graphs across different output formats, the benchmark and the policy bindings designed for each generator are clear methodological contributions. However, the paper does not provide evaluations by architects, data on user behavior, circulation simulations, or other independent measures of architectural performance. Its discussion of applying the method to other design fields also remains primarily conceptual. The paper therefore needs to explain why the difference in integration statistics between generated and real floorplans constitutes a design problem worth addressing, and what practical design benefits these methods provide beyond improving IPD and PIA. Independent validation of design quality, or an analysis of the relationship between changes in integration and actual usage outcomes, would help support the paper’s stronger conclusions.

Questions for authors

Please try to address my questions and concerns in the Critical weakness section.

6 comments
AuthorsOpenReview ↗

Response to Reviewer TnoW: summary

We thank you for the thoughtful review and the recognition of SSIO, the benchmark and the policy bindings. This comment summarizes our responses; the follow-up comments give the details and evidence.

W1 & W2: Do integration-metric gains indicate design improvement, and why is the gap worth addressing?

Concern: The reward, IPD and PIA all derive from SSIO's integration scores, so the experiments mainly show proxy improvements; internal robustness checks do not resolve external validity, and independent evidence of design quality, or a link to use, is needed.

  1. Evidence independent of SSIO. We agree that SSIO-derived metrics cannot by themselves establish design quality. We added evidence that does not use SSIO scores: two assessors outside the project, a real-estate designer and an architecture student, compared 50 condition-matched Base/SSPT pairs blind to model and scores and judged everyday residential use. Both preferred SSPT in most definite choices (83% and 67%), and their recorded reasons concern use. They also mostly preferred the higher-PIA layout (73% and 67%), and a separate author-involved discussion review shows the same pattern (§5.4, Table 5; Appendix D.2.1). Native metrics computed without SSIO are reported with their costs (Table 2).
  2. Why the gap matters. Public spaces lead integration across four residential corpora from 15 countries (Figure 3), and space syntax theory treats the everyday shared space as the integrating core of the home (Hillier, 1996). HouseDiffusion keeps this ordering, but its Public margin is 46.3% weaker (Figure 4). In 12.1% of Base outputs another room, often a bathroom or kitchen, is more integrated than Living, against 1.0% of real homes and 1.9% after SSPT (derived from Table 22). We replace median PIA as a main metric with PIAD, the distance to the real-residence distribution of Living advantages. It complements IPD's typical four-domain profile; a higher PIA is not itself the target (Eq. (2); Appendix B.3).
  3. Scope. Connected-only comparisons, four centrality definitions and a 200-pair case analysis show robustness and what changes in the layouts (Appendix D.1.2, Table 25, Table 26). They do not measure behavior, and we state that inhabitants' use is beyond our scope (§7).

W2 (cross-domain part): The discussion is mainly conceptual

Concern: The cross-domain discussion does not yet establish benefits in other design tasks.

  • We agree and present it as a discussion with a proposed workflow (§6; Appendix F and Appendix G), illustrated by three worked examples without generator training; validation in other domains is listed as future work.

Follow-up comments, in order: W1 & W2 (1/2) and (2/2), then the cross-domain part of W2. Q1 refers to the weaknesses and is covered by them.

AuthorsOpenReview ↗

W1 & W2 (1/2): Evidence independent of SSIO scores

We agree that metrics derived from the same SSIO graphs as the reward cannot by themselves establish design quality, and that internal robustness checks do not resolve external validity. This comment gives the evidence that does not use SSIO scores; the next explains why the integration gap is a design problem and how we scope our claims.

1. Blind assessments by design-trained assessors

We asked assessors with design training rather than lay users. A real-estate designer and an architecture student, neither involved in the project, independently compared the same 50 condition-matched Base/SSPT pairs. Model identities and scores were hidden; the question was which layout better supported everyday residential use. The sample was stratified across test blocks and PIA changes. Both independent assessors preferred SSPT and higher-PIA layouts in most definite choices. Architecture researchers also conducted a separate author-involved discussion review (§5.4, Table 5, p. 9; Appendix D.2.1, Table 27 and Figure 14, pp. 43–45):

AssessmentPrefer SSPTPrefer higher PIABase has higher PIA when Base is preferred
Real-estate designer · independent blind83% (34/41)73% (30/41)43% (3/7)
Architecture student · independent blind67% (28/42)67% (28/42)43% (6/14)
Architecture researchers · author-involved discussion74% (29/39)72% (28/39)40% (4/10)

The first two percentages use definite preferences as their denominator; the last uses Base-preferred cases. These are three assessments of the same 50 pairs. A designer described one SSPT layout as placing the kitchen and bathroom more appropriately relative to bedrooms, enabled by the living-space configuration. Thus, the judgments connect the measured changes to broader use considerations. The recorded examples and procedure are documented in Appendix D.2.1 (pp. 43–44).

2. Native metrics computed without SSIO

HouseDiffusion's own metrics, which do not use SSIO, are reported with their costs on the independent final test (Table 2; acceptance in §5.2):

Metric (final test)Base HD+SSPTMeaning
Compatibility ↓2.5372.572missing or extra access relations vs. the requested graph
FID ↓7.7989.087visual distance of renderings to real plans
Exact-8 acceptance ↑94.5%96.5%measurable outputs that keep all eight spaces
  • Human judgments where conformity worsens. Among the 50 blind-assessed pairs, Compatibility increases for 22. In this subset the real-estate designer still preferred SSPT in 15 of 18 definite choices and the architecture student in 12 of 17 (Appendix D.2.2, Table 28).
  • What these layouts look like. In three examined pairs, the Compatibility increase comes from a one-unit gap between a door and a required room, which the native overlap rule reads as a missing link; this explains a specific count change rather than judging architectural acceptability (Figure 15). Three pairs with higher Compatibility that both independent assessors preferred illustrate functional arrangement, overall layout, and a local overlap change (Figure 16).
AuthorsOpenReview ↗

W1 & W2 (2/2): Why the integration gap is a design problem, and the scope of our claims

1. Residential evidence

Across four residential corpora spanning 15 countries, public spaces have the highest integration (Figure 3, p. 4). HouseDiffusion preserves this ordering, but its Public margin is 46.3% weaker (Figure 4, p. 5). In 12.1% of Base outputs another room, often a bathroom or kitchen, is more integrated than Living, against 1.0% of real homes and 1.9% after SSPT (derived from Table 22, p. 41). In such layouts a bathroom or kitchen, rather than the shared living space, is the most accessible space of the home, as in case B of Figure 7 (p. 8), where SSPT replaces bathroom dominance with Living anchoring under the same eight-room program.

2. Theoretical grounding

The public-space lead also has a theoretical basis. Space syntax theory holds that everyday living belongs in integrated space, where the movement a layout generates is natural to that function, and treats departures from this pattern as possible failures of design (Hillier, 1996, p. 304). The everyday shared space holds the home together: the other functional zones depend on it for their integration, and removing it reduces the layout to a single sequence of rooms (Hillier, 1996, pp. 254, 261). Residential studies likewise report living-room-centered configurations and the private and public aspects of living-room experience (Byun & Choi, 2016; Rechavi, 2009). PIA quantifies this lead per plan (Appendix B.3).

3. What the metrics target

The benchmark compares distributions rather than rewarding a higher PIA, and its two measures divide the work. IPD checks whether the typical integration profile of the four functional domains matches real homes. PIAD, the Wasserstein-1 distance between a generator's per-plan PIA distribution and that of real RPLAN homes, checks plan by plan whether the shared living space leads as it does in real homes: it is 0 only when the distributions coincide and grows when the lead is too weak or too strong (Eq. (2), Appendix B.3). A higher PIA is therefore not itself the target.

4. What the internal evidence shows, and the scope of our claims

Matched conditions, connected-only comparisons, and alternative centralities test internal robustness: gains persist on the 3,914 conditions whose graphs are connected both before and after post-training (Appendix D.1.2, p. 42) and under all four centrality definitions (Table 25, p. 42). As you note, they share SSIO's source. The 200-pair case analysis shows what changes in the layouts: among 144 PIA-increasing pairs, 131 keep the same number of recovered room-level links or fewer, so the gains mainly reflect reorganized access rather than added links (Appendix D.1.5, Table 26, p. 43). We do not claim effects on inhabitants' behavior: the human judgments are expert assessments of layouts rather than observations of use, and the conclusion states that understanding users' intentions and simulated user feedback remain beyond this scope (§7, p. 10). Our claim is accordingly that SSPT moves the spatial organization of generated layouts closer to real residences, and that independent assessors mostly prefer the resulting layouts.

Reviewer_KdKLOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces architectural space syntax into floor-plan generation. The authors propose SSIO, which converts floor plans in different formats into convex-space-based accessibility graphs and computes spatial integration. They then develop two evaluation metrics, IPD and PIA, using RPLAN, and employ these metrics as SSPT-PPO rewards for diffusion model training. The authors report that post-training HouseDiffusion reduces IPD by 27% and improves median PIA by 41.9%. They also apply the reward mechanism to five diffusion models. The central idea is to directly optimize architectural spatial organization rather than relying solely on image similarity.

Critical strengths
  1. The authors develop a unified measurement pipeline for floor plans with different output formats and compare spatial integration across multiple residential floor-plan datasets. Such a measurement tool that can be shared across models has practical value.

  2. Among 4,588 paired results under identical conditions, 70.42% show improved PIA performance for HouseDiffusion. The authors also report bootstrap confidence intervals. Furthermore, the same reward mechanism is applied to multiple diffusion models, demonstrating its potential generality.

Critical weaknesses
  1. SSIO first rasterizes floor plans and then partitions them into convex spaces using maximum inscribed rectangles, relying primarily on door locations to determine connectivity between rooms. This process may produce incorrect connectivity graphs for open-plan living and dining areas or floor plans with missing door information. Although the door-pair F1 score of the GSDiff door-completion model reaches 0.8879, the median PIA error remains 0.3454 (Table 22). Moreover, simply changing the door-completion strategy for the same floor plan can change its Base PIA from 0.5882 to 0.2781 (Table 23). Since both the training rewards and final evaluation depend on these graphs, the authors should independently validate a subset of samples using architect annotations or established architectural floor-plan analysis tools.

  2. In Table 2, after post-training, HouseDiffusion's FID increases from 7.798 to 9.087, while Compatibility increases from 2.537 to 2.572. Although both metrics remain below the authors' predefined thresholds of 9.5 and 2.8, respectively, the results still indicate a decline in generation quality. How much of the improvement in spatial design metrics would be retained under stricter quality constraints that keep performance closer to the baseline?

  3. How SSIO reconstructs the connectivity graph from a floor plan is not sufficiently clear. A step-by-step illustration of rasterization, convex-space decomposition, and door-based connectivity inference would improve readability.

  4. The noise-to-floor-plan illustration in Fig.5 may not reflect the actual generation process. Does the model progressively denoise a triangular-based representation before producing the floor plan? If so, the authors should clarify this process.

Questions for authors

see weakness.

6 comments
AuthorsOpenReview ↗

Response to Reviewer KdKL: summary

We thank you for the detailed review. This comment summarizes our responses to all points; the follow-up comments give the details and evidence.

W1: Are SSIO's access graphs reliable?

Concern: Open living–dining areas or missing doors could yield wrong graphs; the GSDiff completer reaches door-pair F1 0.8879 but a median PIA error of 0.3454, and changing the completion strategy moves Base PIA from 0.5882 to 0.2781. Since rewards and evaluation depend on these graphs, a subset should be validated independently.

  1. Plans with doors (RPLAN, HouseDiffusion). SSIO reads access from explicit doors instead of inferring it. In all 80,788 RPLAN plans, distinct rooms never touch directly, so every annotated passage is an opening. RPLAN has no separate open-plan label: a combined living–dining area is one Living room or two rooms joined by an opening (1,089 of 1,312 Living–Dining pairs). A missing door is a rare annotation gap that we keep as source-data noise; its detectable form, a Dining room without any opening, occurs in 0.08% of plans (Appendix A.2).
  2. GSDiff (doors must be predicted). The numbers you cite concern this case, and we clarify three points. First, we corrected a room-type coding error in that pilot, which changes 0.3454 to 0.3611 (Table 31). Second, the pilot redrew predicted doors into the raster and re-decomposed the rooms; most of the error comes from this redraw step, since redrawing even the annotated doors gives a similar error. The route we deploy keeps the decomposition fixed and inserts access edges, with a median error of 0.0724 on the same 500 held-out plans (Table 32). Third, the change from 0.5882 to 0.2781 is a diagnostic ablation that withholds room types from the predictor, not an alternative completion strategy (Table 33).
  3. Independent validation. RPLAN's plans are real residential designs, and its room and opening annotations come with the dataset, independently of SSIO. The generators we study are trained on RPLAN, and those with doors export layouts in its format, so SSIO measures generated layouts and the real-residence reference with identical rules. We therefore use RPLAN's own annotations as the independent reference for these rules (Appendix A.2, Table 32). This is a reasonable substitute for new labels: in layouts with generated doors, access is explicit rather than inferred, so what needs checking is how SSIO reads doors and openings, and that reading is the same for real and generated plans.
  4. Where design professionals were involved. We agree that expert judgment matters. Because access in these layouts is already given by annotated or generated doors and openings, relabeling edges would largely repeat that information; we therefore asked design-trained assessors for the judgment that genuinely needs their expertise. Two assessors outside the project compared 50 condition-matched Base/SSPT pairs blind to model and scores for everyday residential use, and both preferred SSPT in most definite choices (83% and 67%; §5.4, Table 5).

W2: How much of the gain remains under stricter quality constraints?

Concern: FID and Compatibility worsen despite passing the original limits; would the design gain survive constraints closer to Base?

  • An earlier checkpoint that meets validation FID at most 1.10 times Base and Compatibility at most Base retains 89.8% of the PIAD reduction (95% interval 85.8–94.0%; Table 21). The original limits, 9.5 and 2.8, bound the repeat-to-repeat variation of unchanged Base HD rather than define a tolerated degradation (Table 18). We also show the native-metric changes, blinded judgments for the subset with worse Compatibility, and concrete examples of the counted connection changes (Table 2; Table 28; Figure 15).

W3: Step-by-step illustration of graph construction

Concern: The conversion from a floorplan to the connectivity graph is hard to follow.

  • Figure 8 (Appendix A.2) now illustrates rasterization, convex-space decomposition, door-based connections and the resulting access graph on two generated layouts.

W4: What Figure 5 shows

Concern: Do the triangular shapes depict the actual representation being denoised?

  • HouseDiffusion denoises room-corner coordinates; the triangles come from drawing noisy corners as polygons. Figure 5 shows SSPT for any backbone m; its Native Reverse Generation panel uses HD room-corner denoising as the example, as the revised caption states, and Table 16 specifies the native variables of all five backbones.

Follow-up comments, in order: W1 (1/2) and (2/2), W2, W3, W4. Q1 refers to the weaknesses and is covered by them.

AuthorsOpenReview ↗

W1 (1/2): Openings in RPLAN and the GSDiff numbers

The access source differs across outputs: RPLAN supplies room and opening annotations, HouseDiffusion generates door geometry, and GSDiff omits doors and requires learned access completion. The completion results you cite concern GSDiff. We clarify the opening evidence and evaluate access completion against RPLAN's existing annotations below.

Open plans and missing doors: how RPLAN encodes passages

SSIO builds within-room links from the convex decomposition and links rooms only through encoded openings (Appendix A.2, p. 25; illustrated in Figure 8, p. 26). Whether access is read correctly therefore depends on how the data encode passages, which we audited on all 80,788 RPLAN plans (Appendix A.2, p. 26). Distinct room instances never touch directly (0 of 646,859 adjacent pairs), so every annotated passage is an opening region, and adjacency is not mistaken for access.

  • Open plans. RPLAN has no separate open-plan label. A combined living–dining area is labeled either as one Living room, which SSIO measures as one room, or as Living and Dining rooms joined by an annotated opening: of the 1,312 Living–Dining pairs, 1,089 are joined by an opening, 185 are separated by a wall, and 38 are not adjacent.
  • Missing doors. A passage can be missed only where an opening was not annotated. The detectable case, a Dining room with no opening to any room, occurs in 68 plans (0.08%), including 15 reference plans and 3 benchmark conditions. We keep these plans unchanged as annotation noise in the source data; this count is not an overall graph-error rate.
  • Generated layouts. HouseDiffusion supplies door geometry, which SSIO reads directly; GSDiff omits doors and requires the completion evaluated next.

GSDiff: what the two numbers show

GSDiff outputs room boundaries without doors, so a frozen predictor must complete access before SSIO can measure them (§5.6, p. 10). First, a correction: fixing a room-type coding error in that pilot changes the median PIA error from 0.3454 to 0.3611 (Table 31, formerly Table 22, p. 50). To locate the error, we rescored the same 500 held-out RPLAN plans with annotated and with predicted openings under two access constructions (Appendix E.3.3, Table 32, p. 50); entries are median absolute PIA errors against the original plans:

Access constructionAnnotated openingsPredicted openings
Redraw doors and re-decompose (pilot route)0.36920.3611
Fixed decomposition, inserted access edges (deployed)0.07290.0724

Redrawing even the annotated doors yields a similar error, so the large error comes mainly from re-rasterizing doors and re-decomposing rooms, not from predicting access; GSDiff-SSIO's deployed fixed-decomposition route avoids this step. Prediction still matters within the fixed decomposition (median difference 0.0764 between predicted and annotated access), and these held-out plans have at most seven rooms. Second, the change from 0.5882 to 0.2781 (Table 33, formerly Table 23, p. 51) is a diagnostic ablation that withholds room-type features from the frozen predictor, not an alternative completion strategy; it shows that completed scores depend on room types. Because the completer did not meet our pre-specified per-plan fidelity criteria (3 of 9 met), we report GSDiff-SSIO separately and compare Base with post-training only under the same frozen completion map.

AuthorsOpenReview ↗

W1 (2/2): The independent reference, and where design professionals were involved

1. RPLAN's own annotations as the reference

An independent reference specifies which rooms connect without deriving that answer from SSIO. For real plans, RPLAN provides it: its plans are real residential designs, and its room and opening annotations come with the dataset, independently of SSIO. The corpus audit in the previous comment checks how these annotations encode passages, and the controlled comparison in Table 32 (p. 50) tests completion against them on the same held-out plans. The generators we study are trained on RPLAN, and those with doors export layouts in its format, so SSIO reads generated layouts and the reference with identical rules; for HouseDiffusion, the reward and the evaluation read the doors the model generated, without a prediction step. We therefore use RPLAN's own annotations as the independent reference for the reading rules; because access in layouts with generated doors is explicit rather than inferred, checking the reading rules on annotated plans is a reasonable substitute for labeling generated layouts. Four centrality definitions further test scoring sensitivity (Appendix D.1.4, Table 25, p. 42), and GSDiff-SSIO's Base/post-training comparison uses the same frozen completion map.

2. Where we involved design professionals

We appreciate the suggestion to involve architects, and we agree that expert judgment matters. Access edges are already given by RPLAN's annotations for real plans and by the generated doors in layouts that include them, so relabeling them would largely repeat that information. We therefore asked design-trained assessors for the judgment that genuinely needs their expertise: whether a generated layout suits everyday residential use.

We asked assessors with design training rather than lay users. A real-estate designer and an architecture student, neither involved in the project, independently compared the same 50 condition-matched Base/SSPT pairs. Model identities and scores were hidden; the question was which layout better supported everyday residential use. The sample was stratified across test blocks and PIA changes. Both independent assessors preferred SSPT and higher-PIA layouts in most definite choices. Architecture researchers also conducted a separate author-involved discussion review (§5.4, Table 5, p. 9; Appendix D.2.1, Table 27 and Figure 14, pp. 43–45):

AssessmentPrefer SSPTPrefer higher PIABase has higher PIA when Base is preferred
Real-estate designer · independent blind83% (34/41)73% (30/41)43% (3/7)
Architecture student · independent blind67% (28/42)67% (28/42)43% (6/14)
Architecture researchers · author-involved discussion74% (29/39)72% (28/39)40% (4/10)

The first two percentages use definite preferences as their denominator; the last uses Base-preferred cases. These are three assessments of the same 50 pairs. A designer described one SSPT layout as placing the kitchen and bathroom more appropriately relative to bedrooms, enabled by the living-space configuration. Thus, the judgments connect the measured changes to broader use considerations. The recorded examples and procedure are documented in Appendix D.2.1 (pp. 43–44).

Reviewer_XuChOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper proposes an architectural design-theory-based framework for floorplan generation. It introduces Space Syntax Integration Operator (SSIO) that draws on space syntax architectural theory to convert different floorplan formats into a common intermediate access-graph representation. It also defines the RPLAN-SSI benchmark for evaluating functional organization in residential layouts by building on the RPLAN benchmark. It uses these measures in SSPT, an RL method that fine-tunes diffusion models toward better spatial integration and room hierarchy using an SSIO-based reward.

Critical strengths
  1. I like how the paper connects model-based floorplan generation to established architectural theory, looking beyond the basic validity of layouts in this task.

  2. The experiments and ablations are extensive (but focus on the proposed metrics - see weaknesses).

Critical weaknesses
  1. The evaluation focuses heavily on the proposed metrics, but there is limited evidence that the measured gains correspond to broader and more subjective human-perceived design quality. Additionally the metrics proposed require further explanation and justification. See 'questions to authors.'

  2. The proposed RL-based post-training approach might lead the model to generate floorplans that perform well on the proposed metrics but not on validity metrics that measure things like conformity to the design constraints and organization. This analysis is important and has been under-explored.

Questions for authors
  1. IPD metric makes sense, but I'm confused about the PIA metric. Why does it make sense as a main metric to judge floorplan generators? It seems like something that quantifies a specific, subjective architectural design language. What makes a floorplan with a higher PIA better?

  2. Please have a table of notations and abbreviations as it is difficult for the reader to keep track.

7 comments
AuthorsOpenReview ↗

Response to Reviewer XuCh: summary

We thank you for the constructive review. This comment summarizes our responses to all points; the follow-up comments give the details and evidence.

W1 & Q1: Do metric gains reflect perceived design quality, and why is PIA a main metric?

Concern: The evaluation relies on our own metrics, and PIA may encode a specific design language.

  1. Metric revision. PIA has no absolute optimum: increasing its median can reward an exaggerated Living lead and cannot establish which generator better matches residential organization. We therefore revised the benchmark. In the submission, median PIA was a main metric; in the revision, PIA is an interpretive per-plan quantity, the Living category's normalized integration lead over its strongest competitor. We introduce PIAD, the Wasserstein-1 distance between a generator's per-plan PIA distribution and that of real RPLAN homes, as the main measure of public-space advantage. PIAD replaces median PIA as a main reported metric alongside the existing IPD (Table 1, Table 2; definitions in Eq. (2) and Appendix B.3). PIAD is zero when the distributions coincide; lower means closer alignment. The two measures divide the work: IPD checks whether the typical integration profile of the four functional domains matches real homes, while PIAD checks plan by plan whether the shared living space leads as it does in real homes, so it also registers plans where Living fails to lead or leads too strongly, which domain medians average out. They are correlated but not redundant, ordering three model pairs differently (Table 13).
  2. Human judgments. Two design-trained assessors outside the project, a real-estate designer and an architecture student, compared 50 condition-matched Base/SSPT pairs blind to model and scores. Both preferred SSPT in most definite choices (83% and 67%), and mostly the higher-PIA layout (73% and 67%); a separate author-involved discussion review shows the same pattern (§5.4, Table 5).
  3. Why the public-space lead matters. Public spaces lead integration across four residential corpora from 15 countries (Figure 3), and space syntax theory treats the everyday shared space as the integrating core of the home (Hillier, 1996). This grounds the attribute being measured, without requiring a universally maximal Living lead.
  4. Revision. §3.2 and Eq. (2) define the roles of PIAD and IPD; Table 1, Table 2 and Figure 6 report PIAD; §5.4 and Appendix D.2 report the judgments.

W2: Does post-training trade validity for metric gains?

Concern: RL could improve our metrics while harming validity, such as conformity to the design constraints.

  1. Measured cost. With HouseDiffusion's own metrics, which do not use SSIO, Compatibility changes from 2.537 to 2.572 and FID from 7.798 to 9.087 (Table 2).
  2. Stricter budget. An earlier checkpoint with validation FID at most 1.10 times Base and Compatibility at most Base retains 89.8% of the PIAD reduction (Table 21).
  3. Where Compatibility worsens. Both blind assessors still mostly prefer SSPT (Table 28); in the three examined pairs, the increase comes from a one-unit gap between a door and a room (Figure 15).
  4. Shortcut control. Layout expansion, the shortcut seen during development, is controlled by AreaGuard (Figure 10); the living room is not enlarged beyond real homes (39.9% vs. 40.0% of retained-atom area; Appendix D.1.1).

Q2: Table of notation and abbreviations

Concern: The notation and abbreviations are difficult to keep track of.

  • Added in Appendix A.1 (Tables 7–9) and referenced at the start of §3.

Follow-up comments, in order: W1 & Q1 (1/2) and (2/2), W2, Q2.

AuthorsOpenReview ↗

W1 & Q1 (1/2): What PIA measures and how the benchmark now uses it

Your question led us to revise the metrics. This comment explains why per-plan PIA cannot rank generators, what PIAD measures instead, why we keep both PIAD and IPD, and what the main tables now report. The next comment gives the residential evidence, the blind assessments, and the theoretical grounding (W1).

1. Why PIA is not a benchmark target

PIA is a per-plan quantity: the Living category's normalized integration lead over its strongest eligible competitor, positive when Living leads (Appendix B.3). It explains a layout, but it cannot rank generators, because it has no absolute optimum. Real homes show a typical lead with a spread, not a maximum (RPLAN median 0.6172, interquartile range 0.4838–0.7794; Table 13), so a higher median PIA would also reward an exaggerated living-room lead and would hide how many layouts lack a Living lead at all.

Reading a higher median PIA as better therefore holds only while a generator’s median stays below that of real homes, a condition PIAD does not need. All generators in our experiments meet this condition, and on these results the two summaries reach the same conclusions: the benchmark order is the same, and all five backbones still improve (see item 4 of our “Summary of revisions” comment).

2. PIAD: the main measure of public-space advantage

What PIA probes remains the target: whether the shared living space leads the home's integration as it does in real residences. We therefore compare distributions. PIAD is the Wasserstein-1 distance between a generator's per-plan PIA distribution and that of 20,348 real RPLAN homes (Eq. (2); Eq. (B.1) in Appendix B.3). It is 0 only when the two distributions coincide and grows with any departure in strength or spread, including leads stronger than in real homes, so lower is better and the reference itself scores 0.

3. Why we report both PIAD and IPD

The two measures ask different questions: IPD, whether the typical home has the right functional profile; PIAD, whether each home's shared living space holds the lead it has in real homes. IPD compares the typical relative-integration profile across the Public, Private, Service and Outdoor domains; PIAD compares the full distribution of the within-plan Living lead. IPD's domain medians can hide differences in the tails of that lead, and PIAD does not see how the remaining functions are arranged. Across the eight benchmark systems the two are correlated (Pearson r = 0.93, Spearman ρ = 0.86) but order three pairs differently. FloorplanGAN, for example, has lower IPD than Graph2Plan (0.0943 vs. 0.1078) but higher PIAD (0.4340 vs. 0.3974): Living fails to lead in 28.5% of its layouts, against 8.3% for Graph2Plan (Table 13).

4. The revised main tables

PIAD replaces median PIA in the main benchmark (Table 1, p. 7):

MethodIPD ↓PIAD ↓Acc. (%)
CE2EPlan0.19140.581666.5
DiffPlanner0.18030.554859.1
House-GAN++0.16120.528840.7
House-GAN0.13840.413636.9
Graph2Plan0.10780.397471.6
FloorplanGAN0.09430.434091.5
ChatHD0.08660.351731.2
HD0.11110.412994.5

PIAD also replaces median PIA in the post-training results across five backbones (Table 6, p. 10; HD also in Table 2):

BackboneIPD: Base → +SSPTPIAD: Base → +SSPT
HD0.1111 → 0.08110.4129 → 0.2798
CE2EPlan0.1914 → 0.13520.5816 → 0.4460
DiffPlanner0.1803 → 0.15550.5548 → 0.5043
ChatHD0.0866 → 0.08380.3517 → 0.3446
GSDiff-SSIO0.0386 → 0.03600.0903 → 0.0769

Median PIA and the share of Living-led layouts remain as interpretive columns (Table 13, p. 29).

AuthorsOpenReview ↗

W1 & Q1 (2/2): Residential evidence, blind assessments, and theoretical grounding

1. Residential evidence

Across four residential corpora spanning 15 countries, public spaces have the highest integration (Figure 3, p. 4). HouseDiffusion preserves this ordering, but its Public margin is 46.3% weaker (Figure 4, p. 5). In 12.1% of Base outputs another room, often a bathroom or kitchen, is more integrated than Living, against 1.0% of real homes and 1.9% after SSPT (derived from Table 22, p. 41).

2. Blind assessments by design-trained assessors

We asked assessors with design training rather than lay users. A real-estate designer and an architecture student, neither involved in the project, independently compared the same 50 condition-matched Base/SSPT pairs. Model identities and scores were hidden; the question was which layout better supported everyday residential use. The sample was stratified across test blocks and PIA changes. Both independent assessors preferred SSPT and higher-PIA layouts in most definite choices. Architecture researchers also conducted a separate author-involved discussion review (§5.4, Table 5, p. 9; Appendix D.2.1, Table 27 and Figure 14, pp. 43–45):

AssessmentPrefer SSPTPrefer higher PIABase has higher PIA when Base is preferred
Real-estate designer · independent blind83% (34/41)73% (30/41)43% (3/7)
Architecture student · independent blind67% (28/42)67% (28/42)43% (6/14)
Architecture researchers · author-involved discussion74% (29/39)72% (28/39)40% (4/10)

The first two percentages use definite preferences as their denominator; the last uses Base-preferred cases. These are three assessments of the same 50 pairs. A designer described one SSPT layout as placing the kitchen and bathroom more appropriately relative to bedrooms, enabled by the living-space configuration. Thus, the judgments connect the measured changes to broader use considerations. The recorded examples and procedure are documented in Appendix D.2.1 (pp. 43–44).

3. Theoretical grounding

The public-space lead also has a theoretical basis. Space syntax theory holds that everyday living belongs in integrated space, where the movement a layout generates is natural to that function, and treats departures from this pattern as possible failures of design (Hillier, 1996, p. 304). The everyday shared space holds the home together: the other functional zones depend on it for their integration, and removing it reduces the layout to a single sequence of rooms (Hillier, 1996, pp. 254, 261). Residential studies likewise report living-room-centered configurations and the private and public aspects of living-room experience (Byun & Choi, 2016; Rechavi, 2009). PIA quantifies this lead per plan (Appendix B.3).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions (to all reviewers and the AC)

We thank all reviewers for their careful reading and constructive suggestions. We have revised the paper accordingly; the main changes are listed below, and our responses to each review give the details. The method, trained models, checkpoint selection and evaluation protocol are as submitted, and no model was retrained.

Presentation

The presentation scores indicated that parts of the paper were hard to follow. Beyond the specific points raised in the reviews, we therefore revised the presentation throughout:

  1. We rewrote passages that were hard to follow or ambiguous, aiming for shorter and more direct explanations, and added a reference of terms, notation and abbreviations (Appendix A.1, Tables 7–9).
  2. We added a step-by-step illustration of how SSIO turns a floorplan into an access graph (Figure 8, Appendix A.2) and clarified the HouseDiffusion schematic in Figure 5.
  3. We corrected imprecise statements, including the GSDiff pilot's median PIA error (0.3454 → 0.3611 after fixing a room-type coding error; Table 31) and the meaning of the quality limits (Appendix C.4).

Addressing concerns and strengthening evidence

  1. Metrics: PIAD, the distance between the per-plan PIA distributions of generated and real homes, is now the main measure of public-space advantage, paired with IPD for the overall functional profile; per-plan PIA and its median are interpretive quantities (§3.2, Eq. (2)).
    • What changed. Only how per-plan PIA is summarized for each generator. Reviewer XuCh asked what makes a higher PIA better (Q1). The submission marked median PIA as higher-is-better, but PIA has no absolute optimum: above the level of real homes, a higher median would reward an exaggerated Living lead. PIAD therefore compares the whole per-plan PIA distribution with that of real homes, so the target is the reference itself.
    • Why the two agree on our results. Every generator lies below the reference in median PIA, and the benchmark generators lie largely below it: for each, even the 75th percentile is below RPLAN's median of 0.6172 (Table 13). Moving such a distribution toward real homes raises its median and lowers PIAD together, and when a distribution is below the reference at every quantile, PIAD equals its gap in mean PIA. This is also why Figure 6(b) of the submission could measure PIA gains as closure of the gap to RPLAN's median.
    • Effect on the reported results. The eight benchmark generators have the same order under both summaries, and all five backbones improve under each. In the three ablations, the only change in order is a near-tie between no area penalty and AreaGuard-2 (submission median PIA 0.3912 vs. 0.3921; PIAD 0.2855 vs. 0.2890), which leaves that ablation's conclusion unchanged. Median PIA remains reported for the benchmark and the main result (Table 13, §5.2).
  2. Human judgments: two independent blind assessors (a real-estate designer and an architecture student) compared 50 condition-matched layout pairs for everyday residential use; an author-involved discussion review is reported separately (§5.4, Table 5; Appendix D.2).
  3. Case analysis: the case study grows from ten illustrative pairs to a 200-pair systematic analysis grouped by recurring themes (Appendix D.1.5, Table 26).
  4. Further rigor: a stricter quality budget (Table 21), a controlled GSDiff access-route comparison against RPLAN's own annotations (Table 32), an audit of how all 80,788 RPLAN plans encode passages (Appendix A.2), and sensitivity to four centrality definitions (Table 25).

The abstract now states the roles of PIAD and IPD and presents the cross-domain part as discussion. We have also updated the anonymous code repository (https://anonymous.4open.science/r/GenAI4Design/README.md) with the PIAD computation, a CPU-only script that recomputes it from score files, and documentation of the revised metric roles.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.