acceptodds
Under review as a conference paper at ICLR 2027

HouseDiffusion++: Constraint-Consistent Vector Floorplan Generation

Abstract

In real architectural design scenarios, floorplans must satisfy a variety of requirements imposed by regulations, design briefs and project-specific constraints. Often combining geometric, functional, and semantic requirements, these conditions make practical floorplan generation a highly constrained design problem rather than an unconstrained layout synthesis problem. Current floorplan generation methods usually address only a small subset of these conditions. In this work, we introduce HouseDiffusion++, a constraint-consistent diffusion framework for controllable vector floorplan generation. Building on HouseDiffusion, our approach extends vector denoising with several conditioning mechanisms to enable joint control over room adjacency graphs, outer boundaries, room sizes, windows and entrance doors. We also address a significant mismatch between training and inference in the original formulation, whereby the number of room corners is known during training but unavailable at test time. To overcome this issue, we introduce a polygon augmentation strategy that enhances robustness to redundant points. Experiments on the RPLAN dataset demonstrate that HouseDiffusion++ outperforms baselines in terms of quantitative metrics and visual quality, while offering greater controllability under different constraints.

Primary area:
generative models
Keywords:
Diffusion Models, Controllable Generation, Floorplan Generation, Structured Generation, Geometric Constraints, Vector Graphics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3Qt4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes a method for floorplan generation conditioned on room adjacency graphs and outer boundaries. It additionally proposes conditioning mechanisms for room sizes and window/door positions. The title and abstract promise a “constraint-consistent diffusion framework” which could be understood as strictly enforced constraints, but they are soft constraints implemented through guidance. Further, the method proposes a simple workaround for a limitation inherited from HouseDiffusion that requires knowing the number of room corners in advance of the floorplan generation.

Critical strengths

None

Critical weaknesses

A key claimed contribution of the paper is the “polygon augmentation strategy”. The baseline HouseDiffusion (CVPR 2023) assumes that the number of per-room corners is known in advance of the diffusion process (from dataset room-type statistics). This is a limitation that the proposed method tries to solve via data augmentation. I.e. it adds redundant points that can then simply be placed on a straight line if not needed. While effective, it does not seem particularly elegant (this is my subjective opinion, of course). Other methods actually learn the number of room corners (see RoomFormer uses a parallel Transformer, SceneScript an auto-regressive Transformer) while this method learns to become more robust to a wrong given number of corners. A more elegant approach would be to learn which corners are indeed needed via an additional classification score. This could be a meaningful baseline, and also result in a more compact vectorized representation getting rid of redundant points.

The paper’s other claimed contribution is the joint conditioning on a number of properties, including the room size and door/window positions. While this work is the first to propose that specific task, the technical implementation is not evaluated clearly. For example, the proposed spatial conditioning represents boundaries by adding 8 additional vertices along each edge, which are then encoded with a “Spatial Encoder” attended to via cross-attention. GSDiff proposes an alternative approach that renders the boundary to a 256x256 image that is then attended to via cross-attention. This work only explains what they do, but there is no comparison to prior approaches -- it would be good to have an analysis that compares different ways of providing the conditions to know which one to choose.

Questions for authors

How would this method compare to a baseline that learns the number of corners instead of doing data-augmentation to become robust towards a wrong initial corner-number?

The work proposes different conditionings, but the actual conditioning mechanisms are not compared to prior work - e.g. the graph and the boundary conditioning. Can such a comparison be provided?

Remarks: l.61 - what are “vector diffusions”? Isn’t every diffusion on a vector? Fig. 2: The comparison of different methods with different conditionings seems somewhat misleading. They should be split into different Figures depending on the conditioning types.

2 comments

Q1. How would this method compare to a baseline that learns the number of corners instead of doing data-augmentation to become robust towards a wrong initial corner-number?

We trained a separate predictor to estimate each room’s corner count before generating the floorplan. We initialized it from the coordinate diffusion model trained without corner augmentation and fine-tuned it for 10 epochs. The predictor uses one token per room, encoding its type, identity and target area. Each room attends to all rooms in the apartment, to its neighbors in the adjacency graph, and to the exterior boundary through cross-attention. A linear head predicts one scalar per room. We trained it with mean squared error against the ground truth corner count, then rounded each prediction to the nearest integer with a minimum of four. We supplied these budgets to the diffusion generator trained without corner augmentation.

Corner-count prediction. On the 600-plan RPLAN split used for predictor validation (4,816 rooms), the predictor achieved a mean absolute error of 0.575 corners per room before rounding and 0.506 after rounding. Always allocating four corners gave an MAE of 1.216.

Floorplan generation. We compare methods on the same data as in Table 8 of the appendix. All rows use room-size conditioning with guidance scale 3.

GeneratorRoom budget at inferenceFID ↓KID ↓Boundary IoU ↑GED ↓Size error ↓Window/door IoU ↑
Without augmentationGround truth4.710.310.9830.210.0170.924
Without augmentationLearned count14.737.500.9560.340.0300.940
Without augmentationExtra vertices16.2311.870.9470.410.0510.926
Ours, with augmentationExtra vertices5.410.920.9840.220.0170.953
Ours, with augmentationN=85.450.940.9820.220.0180.954

Even at half a corner of error per room, replacing the true counts with learned counts increased FID from 4.71 to 14.73 and size error from 0.017 to 0.030. Table 8 of the appendix shows how augmentation changes the generator’s response to surplus vertices. From Ground truth vertices (Original) to Extra vertices (More Points), FID changes from 5.38 to 5.41 with augmentation and from 4.71 to 16.23 without it.

When ground-truth corner counts are unavailable at inference, the augmented generator with a fixed budget (N=8N=8) outperforms the generator using learned counts on all six metrics in this experiment. For example, it achieves a FID of 5.45 versus 14.73, a GED of 0.22 versus 0.34, and a size error of 0.018 versus 0.030. The fixed budget also removes the need to train a separate corner-count predictor.

Q2. How does the conditioning mechanism compare with prior work?

We did not ablate the graph-conditioning mechanism because graph-guided attention is central to the original HouseDiffusion architecture, which we use as our backbone. Replacing it would change how the backbone models room relationships. We therefore retain this mechanism as designed and focus our ablation on the spatial conditioning introduced in this work.

We compared three ways to give the denoiser the apartment geometry. In our original model, a spatial encoder represents each boundary segment as a token. Every floorplan vertex can attend to these tokens through a dedicated spatial room cross-attention layer (SRCA). Our first alternative uses the same encoded segments but appends the boundary tokens to the sequence of floorplan vertices. Global self-attention then lets the vertices and the boundary tokens attend to one another. This changes how the boundary enters the denoiser while keeping the SRCA pathway for windows and the entrance door.

Our second alternative takes the spatial condition as an image. Raster boundary encoding has precedent in Graph2Plan, which uses a CNN. For our diffusion model, we render the boundary and any supplied openings in separate channels of a single 224 × 224 image. A frozen DINOv2 ViT-B/14 encodes the image into 256 patch tokens, and every floorplan vertex can attend to every patch through cross-attention.

We evaluated checkpoints on the same setup as in Table 1 of the main paper. CFG 0 disables target room areas and CFG 3 enables size guidance.

Size CFGSpatial conditioningFID ↓FIDgs ↓KID ↓KIDgs ↓Boundary IoU ↑GED ↓Edge IoU ↑SizeError ↓
0Vector segments + SRCA (ours)6.434.210.890.630.9660.3300.9300.163
0Vector segments + boundary concatenation6.524.331.131.020.9680.2960.9380.160
0Raster + DINOv2 cross-attention6.214.200.640.630.9570.2870.9310.163
3Vector segments + SRCA (ours)6.053.980.960.520.9740.2820.9420.028
3Vector segments + boundary concatenation5.994.150.700.720.9760.2580.9460.021
3Raster + DINOv2 cross-attention6.073.870.740.770.9670.2710.9420.023

The three conditioning mechanisms achieve broadly comparable results on this benchmark. At CFG 3, FID ranges from 5.99 to 6.07 and Boundary IoU from 0.967 to 0.976. Boundary concatenation gives the highest Boundary IoU at both guidance scales, while raster conditioning gives the lowest KID and GED at CFG 0. These results support all three designs as viable ways to condition a vector floorplan generator on spatial constraints.

Q3. Vector diffusions

By “vector diffusions,” we meant diffusion models that directly denoise polygon vertex coordinates. We will use this more precise wording in the final version.

Reviewer_ipa2OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces HouseDiffusion++, a framework builds upon a competitive baseline HouseDiffusion, for conditional floorplan generation. The key contribution involved the support for multi-constrained input conditions such as room boundary, relative room size, together with the original bubble diagram graph input. The paper designs SRCA, a strightforward attention mechanism to encode the boundary constraint, and a transformer architecture to encode the size constraint. The paper further adopts a simple but effective data augmentation approach to mitigate the ambiguity on corner number assignment for each room during inference. Extensive experiments and ablation studies are conducted on RPLAN benchmark dataset. Compared with the previous counterpart conditioned on a single condition, the method has achieved notable improvement.

Critical strengths
  1. The paper is well written and easy to follow, with good demonstration on its approach and clear diagrams.
  2. Authors tackle a challenging yet practical problem, on interior floorplan generation with multi-dimentioned constrains. This setting is less covered by the previous methods where most of them mainly focus on a sole constraint.
  3. The method is strightforward, modular designed and act as a plug-and-play module on top of HouseDiffusion. The data augmentation approach for tackling the ambiguity on corner assignment during inference handles a critical limitation on HouseDiffusion.
  4. The experiments are extensive and the ablation study could verify the motivation and effectiveness of the method.
Critical weaknesses
  1. To me, the critical limitation is the unfair setting when compare with HouseDiffusion, which is the baseline where the method builds upon. The multi-constrained input clearly provides more information and the generated floorplans will benefit from this and stick better to the original distribution. A fairer comparison is to only provide graph inputs to HouseDiffusion++ and compare with HouseDiffusion, or revise HouseDiffusion in a feasible way to encode the size & boundary inputs.
  2. All experiments are done on RPLAN dataset. I've worked extensively on this dataset and I believe the overall style and layout is not diverse enough. This flags the overfitting risk when directly split training / validation / testing on the same dataset. To fully validate the effectiveness / generalizability of the method, I suggest authors to test the model on some other floorplan layout dataset such as Cubicasa, Structured3D, etc.
Questions for authors

I value the paper on its simplicity, effectiveness and practicality. However, I have a few major concerns on the experimental designs as well as the benchmark used by the paper. I am on 'weak accept' at this time, however I encourage authors to address my concern to make the contributions more convincing.

1 comment

We added two experiments in the appendix section “Graph-only comparison and transfer to CubiCasa5K,” together with quantitative results and qualitative examples.

Comparison under equal inference inputs. We evaluated HouseDiffusion++ with only the room graph on the same RPLAN test split as Table 1, using the same ground-truth vertex counts as HouseDiffusion. HouseDiffusion++ achieves lower GED (0.570 vs. 0.760), but higher FID (15.90 vs. 8.42). Our checkpoint received boundary conditions throughout training, so removing them introduces an out-of-distribution conditioning regime. We report these mixed results alongside the boundary- and area-conditioned configurations to clarify the contribution of additional inputs.

Generalization beyond RPLAN. We evaluated both methods on 646 randomly selected CubiCasa5K floorplans using frozen RPLAN checkpoints, without training or fine-tuning on CubiCasa5K. For HouseDiffusion++, providing the boundary improves BoundIoU from 0.365 to 0.794 and reduces FID from 103.85 to 42.34. Adding target room areas further reduces SizeError from 0.590 to 0.371. These results show that spatial and area conditioning remain useful beyond RPLAN.

Reviewer_zuUbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This work extends HouseDiffusion to generate vector floorplans conditioned on building boundaries, room adjacency, room sizes, and predefined openings. It introduces spatial cross-attention, size conditioning, and polygon augmentation for fixed-vertex generation. Experiments on RPLAN, with synthetic window annotations, show improved generation quality and constraint adherence over selected baselines.

Critical strengths

The method directly generates editable vector floorplans rather than raster images, avoiding raster-to-vector conversion and its associated geometric errors. This is valuable for practical architectural design applications.

Critical weaknesses
  1. The method requires users to specify detailed constraints, including room sizes and exact opening locations, which may be difficult and time-consuming in real design scenarios. Meanwhile, it cannot directly constrain individual room boundaries. The paper does not sufficiently demonstrate the practical value of this constraint-setting approach.

  2. Doors and windows are represented as independent polygons rather than openings attached to walls. Even when exterior openings are given as fixed conditions, the model regenerates their coordinates, potentially creating conflicts with the input constraints. Attention masks encode graph relationships but do not guarantee geometric consistency. The need for additional geometric post-processing further raises concerns about the claimed constraint consistency.

Questions for authors
  1. What realistic design workflow motivates requiring users to specify detailed room sizes and exact opening locations? How does this compare with manually creating or editing a floorplan?

  2. Why are predefined exterior doors and windows regenerated rather than kept fixed? How does the model guarantee that generated openings remain correctly attached to walls, particularly when predicted wall positions conflict with the given opening geometry?

  3. What percentages of raw generated floorplans contain self-intersections, overlapping rooms, or uncovered regions? How much does post-processing improve these results?

  4. Why is the edge-sampling parameter L=8L=8 used without an ablation study? How does the method compare with more recent state-of-the-art vector floorplan generation methods under equivalent conditions?

3 comments

Q1. What design workflow motivates specifying room areas and opening locations?

When architects design the facade before partitioning an apartment into rooms, the window locations are already known and guide the placement of interior walls. The building design may also fix the entrance location, and the developer’s brief may specify target room areas. In our model both inputs are optional: a designer can provide just the apartment boundary and room graph and let the model determine room areas and opening locations. This workflow motivated a concrete request from the architects we worked with. They asked us to support target room areas and to preserve window and entrance-door locations established earlier in their design process. When targets are available, conditioning on room areas reduces SizeError from 0.163 to 0.028 (Table 1), and conditioning on opening locations yields a window/door IoU of 0.95 (Table 2).

Area control also supports editing. A designer who likes a layout but wants a larger bedroom can change its target area and ask the model to redistribute the available space, saving the designer from having to reposition each affected partition manually. Figure 3 illustrates how the selected room changes size while the overall room arrangement remains similar.

Q2. Why regenerate predefined openings, and how reliably do they attach to walls?

Regenerating predefined windows and entrance doors allows us to use a single, more general model for both conditional and unconditional opening placement. When opening polygons are supplied (cond), the model learns to reproduce their geometry as it generates the rooms. When they are withheld (uncond), the same model learns to place the openings using the layout-level conditions. Even under this more general formulation, the regenerated polygons closely match the supplied geometry, achieving a window/door IoU of 0.95 in Table 2.

Wall attachment is learned and has no hard guarantee. In practice, however, most generated windows and entrance doors remain attached to the walls of their intended rooms in both modes.

To measure attachment directly, we evaluated the same test plans as in Table 2 of the main paper. We scored the generated polygons directly. An opening counts as attached if it straddles the boundary of its annotated host room. We applied this test to both the ground-truth host room and its generated counterpart. A plan passes when every opening passes.

Opening locationsHost room used for evaluationWindows attachedEntrance doors attachedAll openings attached on plan
Supplied (cond)Ground truth99.00%100.00%97.50%
Supplied (cond)Generated98.17%94.67%90.67%
Withheld (uncond)Generated99.33%92.50%91.33%

With supplied locations, all openings attach to the original room walls on 97.50% of plans and to the generated room walls on 90.67%. With locations withheld, the model can choose a different placement from the reference and still attach the opening to a wall.

Additionally, we assume that the supplied opening locations are geometrically consistent with the apartment boundary. For example, a window or entrance door specified away from the boundary constitutes an inconsistent input, and the model is not designed to resolve such conflicts automatically.

Q3. What percentages of raw generated floorplans contain self-intersections, overlapping rooms, or uncovered regions? How much does post-processing improve these results?

We evaluated floorplans from the Table 1 experiment at guidance scale 3 and applied post-processing to those same predictions. Post-processing eliminated self-intersections on this set and reduced the percentage of plans with overlapping rooms from 19.00% to 0.83%.

Violation and measureRaw plans affectedPost-processed plans affectedRaw mean per planPost-processed mean per plan
Self-intersections (rooms with a self-intersecting boundary)7.33%0%0.075 rooms0 rooms
Overlap (fraction of the room union covered by at least two rooms)19.00%0.83%0.142%0.047%
Uncovered area (affected if the uncovered fraction exceeds that of the paired ground truth)66.17%0.17%2.745%0.0187%

We measured room polygons directly, excluding doors and windows. Each self-intersecting room counts once. Overlap requires positive area, so rooms sharing only a wall contribute zero.

For each plan, we compute the uncovered fraction by dividing the area inside the apartment footprint that no room polygon covers by the total footprint area. We count a generated plan as affected if this fraction exceeds that of its ground-truth annotation. By this criterion, 66.17% of raw plans are affected, yet the mean uncovered fraction is 2.745% for raw predictions versus 2.501% for ground truth, a difference of only 0.244 percentage points. The training samples contain small gaps between room polygons, and these similar averages are consistent with the model learning that spacing. Post-processing reduces the mean uncovered fraction to 0.0187%.

Post-processing can still leave rooms overlapping and uncovered area when the raw placement is poor. For example, if one predicted room lies entirely within another, the smaller room can receive no grid cells of its own and retain its original polygon. However, such cases are rare, and in most cases post-processing helps to resolve artifacts.

Q4. Why is the edge-sampling parameter (L=8L=8) used without an ablation study? How does the method compare with more recent state-of-the-art vector floorplan generation methods under equivalent conditions?

We inherited L=8L=8 from the original HouseDiffusion edge embedding and kept it fixed throughout our experiments. We do not claim this is the optimal choice.

Unfortunately, only few recent vector floorplan generators provide both inference code and pretrained weights. We evaluated DiffPlanner (2025) and CE2EPlan (2026) on the test set from Table 1 using the authors’ checkpoints and post-processing. We report raw outputs for our model, following the protocol in Tables 1 and 5.

All methods receive the apartment boundary and the requested room counts and types. Both baselines also receive the entrance location. DiffPlanner and our model receive the target room-adjacency graph. CE2EPlan uses the requested room types in mode t and additionally receives the ground-truth room centers in mode tl (the center of each room’s bounding box). Neither row for our model uses target room areas. The oracle row uses the annotated vertex count for each room, as in Table 1. The fixed-budget row uses N=8N=8 as in Table 5, allocating eight vertices per room, sixteen to the living room, and four per door.

MethodAdditional room constraintsFID ↓FIDgs ↓KID ↓KIDgs ↓BoundIoU ↑GED ↓GED* ↓Edge IoU* ↑
DiffPlannerGraph20.1010.6417.236.830.9743.082.150.669
CE2EPlan (t)—21.2610.8419.097.190.9744.252.270.645
CE2EPlan (tl)GT centers21.7710.7022.119.840.9753.631.260.831
Ours (oracle vertex counts)Graph6.434.210.890.630.9660.330.100.986
Ours (fixed N=8N=8)Graph6.654.361.030.910.9650.310.110.988

* CE2EPlan and DiffPlanner place interior doors during post-processing, so which room pair each generated door will connect is unknown in advance, while the original GED infers these connections from geometry. For GED* and Edge IoU*, we use the same assignment rule for all methods. In this setting a door can realize a target edge if it touches both rooms, including at their boundaries. We use the ground-truth room graph to match doors to target edges one-to-one, maximizing the number of realized edges. Each remaining door that touches at least two rooms is assigned to the pair with the greatest combined overlap. We count each distinct pair once, including extra pairs. GED* measures the room-type-aware edit distance to the target graph. Edge IoU* measures the intersection over union of the two edge sets.

DiffPlanner provides the closest comparison because both methods use the target graph. With a fixed vertex budget, our model achieves lower FID (6.65 vs. 20.10), KID (1.03 vs. 17.23), and GED (0.31 vs. 3.08). DiffPlanner achieves higher BoundIoU (0.974 vs. 0.965). Both CE2EPlan modes also achieve higher BoundIoU, while our fixed-budget model has lower FID, KID, and GED than either mode.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.