acceptodds
Under review as a conference paper at ICLR 2027

FlowPlan: Joint Structural and Geometric Generation of Polygonal Floorplans via Flow Matching

Abstract

Floorplan generation is an increasingly prominent challenge at the intersection of computer vision, computer graphics, and architecture. Early rule-based heuristics often fail to generalize across diverse, real-world datasets. Recent generative methods have achieved strong performance, but are predominantly evaluated on relatively regular residential layouts and often rely on predefined room semantics, topology, or cardinality. In this work, we introduce FlowPlan, a "generate all" approach that concurrently produces both discrete room semantics and their corresponding continuous geometry. By conditioning the generation process solely on the floor plan boundary, our method eliminates the need for predefined semantic assumptions. We explore various flow matching formulations for discrete variables and evaluate our approach against established baselines on a real-world dataset. Due to our method's flexibility, a single model can be supplied with additional conditioning at test time and thus compared against baselines with and without semantic generation capabilities, demonstrating performance on par with highly specialized models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Flow Matching, Floorplan Generation, Computer Vision, Deep Generative Models

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_URZ3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

FlowPlan employs a single-stage joint continuous-discrete flow matching model to address the problem of generating floor plans based on building footprints. Instead of sampling room diagrams or edge structures before generating geometry, this method predicts corner coordinates, room connectivity, and room type solely based on the vectorized building footprint encoded through a cross-attention mechanism.

Critical strengths

1.This paper employs a clear and concise design: room and corner indices are generated within the same model, thus eliminating the need for a separate edge prediction stage to reconstruct connectivity. Furthermore, the shared backbone network allows for controlled comparisons of continuous, discrete, and order-variable formulas. I believe this is a valuable solution approach.

2.The paper is clearly structured, and experiments effectively demonstrate the effectiveness of the method.

Critical weaknesses
  1. The main baseline is a GSDiff pipeline reconstructed based on a zero-thickness shared edge convention and converted back to the original format with the average wall thickness. This adjustment may help the baseline, so the size of the header gap cannot be entirely attributed to the joint generation, but rather to the choice of transformation and post-processing.

  2. This comparison only includes one baseline model. Given that existing planar graph generation methods, including rule-based methods (e.g., Holodeck [1]), can also partition space based on common sense and user needs, could the authors provide more comparative models to better demonstrate the advantages of the proposed method?

  3. As shown in Section 4.5.4 (Failure Cases), the proposed method seems to be limited by its training data and planar graph configurations. Could the authors further explore the generalization ability of this method to unseen planar graphs and more diverse spatial layouts?

[1] Holodeck: Language Guided Generation of 3D Embodied AI Environments

Questions for authors
  1. This comparison only includes one baseline model. Given that existing planar graph generation methods, including rule-based methods (e.g., Holodeck [1]), can also partition space based on common sense and user needs, could the authors provide more comparative models to better demonstrate the advantages of the proposed method?

  2. As shown in Section 4.5.4 (Failure Cases), the proposed method seems to be limited by its training data and planar graph configurations. Could the authors further explore the generalization ability of this method to unseen planar graphs and more diverse spatial layouts?

3.In the experiment, I want to investigate whether learning-based methods have a speed advantage, since the reasoning process shown in the teaser is step-by-step (200 steps in total), which means inference time could be a disadvantage. I'd like to see some concrete comparative results; could you explain and demonstrate that?

2 comments

We thank the reviewer for their constructive feedback. Before we proceed to address the questions we would like to report new improved results, achieved by performing stochastic inference using the same trained model, i.e. this is an inference time adjustment. We observe that injecting noise during the inference of continuous quantities (coordinates) substantially improves both quantitative and qualitative results. Our intuition is that this noise injection helps the model recover from sub-optimal trajectories in under-explored regions. Specifically, instead of performing a standard Euler step:

xtn+1=xtn+vθ∆tx_{t_{n+1}}= x_{t_n} + v_θ ∆t,

we adopt a two-step procedure, referring to this as Renoising:

  • we estimate the clean data sample: $ \hat x_1 = x_{t_n} + v_θ (1 −t_n) $
  • we perform a forward noising step at tn+1t_{n+1}: xtn+1=(1−tn+1)ε+tn+1x^1x_{t_{n+1}} = (1-t_{n+1})ε + t_{n+1}\hat x_1

Furthermore, regarding the discrete quantities, we found that applying corrector sampling[1], which increases the probability of re-masking a token at each step, also yields consistent improvements. Quantitative and qualitative ablation results on the MagicPlan dataset are reported in the table below and in Figure 12 in Appendix A.2

MagicPlan Dataset | Inference settings ablation

Model NameFID↓FID-Dinov2↓FID-Geom.↓IoU↑
OE-DFM6.7594.1126.860.94
OE-DFM w/ Renoising3.4131.4518.800.96
OE-DFM w/ Corrector Sampling6.6092.1826.440.94
OE-DFM w/ Renoising & Corrector Sampling3.0830.4918.300.96

[1]Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky T. Q. Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete flow matching. In Proceedings of the 38th International Conference on Neural Information Processing Systems, volume 37 of NIPS ’24, pp. 133345–133385, Red Hook, NY, USA, December 2024. Curran Associates Inc. ISBN 979-8-3313-1438-5.

Here, we are addressing the concerns of the reviewer.

Additional baseline and dataset

The reviewer asks for additional baseline models to further demonstrate the performance of our approach. We provide new results comparing our method against the recent state-of-the-art autoregressive approach FMLM on the MagicPlan dataset and also report further results for all the models on the RPLAN dataset. Notably, our reproduced results for GSDiff on RPLAN closely align with those reported in the FMLM paper. Furthermore, like Holodeck, FMLM is a language modeling-based approach. However, we select FMLM for our comparison because it is explicitly designed for floorplan generation and represents the current state-of-the-art performance in this domain.

MagicPlan Dataset

ModelType#Params (in millions)FIDFID-GeomFID-DinoIoU
GSDiffParallel13712.4927.82138.200.88
FMLMAutoRegressive1354.3114.6744.680.90
OurParallel1443.0818.3030.480.96

RPLAN Dataset

ModelType#Params (in millions)FIDFID-GeomFID-DinoIoU
GSDiffParallel13712.0510.6092.250.95
FMLMAutoRegressive1351.2411.5742.680.96
OursParallel1441.375.7116.520.99

Effect of preprocessing and postprocessing pipeline on GSDiff performance

The reviewer challenges the effect of the preprocessing and postprocessing pipeline on the performance of GSDiff. To address this we also report results on RPLAN. Because GSDiff requires no preprocessing or postprocessing on RPLAN, the results more accurately reflect its performance. Moreover, the performance ranking of the models remains identical across both datasets, suggesting that our processing pipeline is a valid proxy for datasets with wall gaps.

Generalization ability of our approach

The reviewer requests experiments on unseen footprints to evaluate the generalization ability of our approach. In Appendix D, we present cross-dataset generalization experiments between our two evaluated datasets, performing inference on footprints from the unseen dataset. It is important to note that RPLAN contains simpler, Manhattan-type building footprints with fewer vertices on average compared to MagicPlan (13 vs. 17). We observe that the model trained on RPLAN struggles to follow non-Manhattan building footprints, included on MagicPlan, and exhibits difficulties with complex boundaries. Conversely, the model trained on the more diverse MagicPlan dataset generates in most of the cases valid floorplans when conditioned on the simpler Manhattan-type footprints of RPLAN. This further validates our adoption of MagicPlan as our main benchmark.

Inference speed

The reviewer is concerned about the inference speed. Compared to other learning-based methods we achieve fast inference independently of the dataset. On MagicPlan we report inference times per floorplan in ms:

  • GSDiff: 909.54
  • FMLM: 760.23
  • Ours: 249.68

On RPLAN we report:

  • GSDiff: 729.20
  • FMLM: 202.76
  • Ours: 236.60

For FMLM, the difference in speed is caused in part from the fact that MagicPlan requires on average more tokens to describe a sample floorplan, requiring 162 tokens in contrast to 126 tokens required for an RPLAN sample.

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

FlowPlan generates polygonal floorplans from building footprints by jointly modeling continuous corner coordinates and discrete room attributes through flow matching. It also supports additional semantic and structural conditioning through test-time inpainting.

Critical strengths

The unified representation of geometry, room semantics, and variable room structures is interesting and flexible. Inpainting enables additional conditioning without retraining. I also learned a lot from the discrete flow matching section, which I found very interesting and useful.

Critical weaknesses
  1. Incomplete evaluation across datasets and baselines. I compared the boundary-constrained generation results in the original GSDiff paper with the GSDiff results in Figures 6 and 8 of this paper. The visual results in this paper is a little bit weaker to me. Dataset differences and the additional preprocessing may contribute to this discrepancy, but their effects remain unclear. It would be helpful to provide tables and visualizations on RPLAN dataset using FlowPlan, GSDiff, WallPlan, Graph2Plan, or other relevant baselines.

  2. Geometric validity is insufficiently evaluated. In my understanding, the representation itself does not automatically guarantee valid room polygons. For example,

  • Two-vertex rooms. When corner counts are not specified, a room may be assigned only two active vertices, which cannot form a valid polygon.
  • Room overlap. The same area may belong to two or more rooms.
  • Boundary violations. Predicted vertices may lie outside the building boundary. Even when all vertices are inside, the edges connecting them may still be outside the boundary.
  • Uncovered regions or holes. Some regions may not be covered by any room. Although gaps can represent walls, larger holes may remain that cannot reasonably be explained by wall thickness.
  • Self-intersection. Suppose the correct boundary order of a rectangle’s corners is 1–2–3–4. A predicted order of 1–3–2–4 would produce a self-intersecting quadrilateral.

Some of these failures may theoretically occur in both FlowPlan and GSDiff. The reported FID, FID-DINOv2, FID-Geom are useful, but they do not fully measure these geometric errors. Comparisons of geometric validity across FlowPlan, GSDiff, or MHD, together with an explanation of how invalid outputs are handled, would therefore be helpful.

  1. The evidence for competitive graph-conditioned generation is limited. The FlowPlan support room-graph condition by inpainting, meaning freezing room attributes while denoising coordinates. The authors state, “We show competitive performance even though our model was never trained for this setting” in lines 353–354. However, Table 2 shows that MHD w/o Adj. outperforms OE-DFM on all four metrics and DFM on all three FID metrics. Supporting this setting without retraining is useful, but the source of the performance gap remains unclear. In my understanding, your inpainting setting is also something could be trained, by partial masking in flow matching training. After inpainting training, would the gap get narrowed with MHD w/o Adj.?
Questions for authors
  1. Could you provide quantitative and qualitative comparisons on RPLAN? Comparisons with GSDiff, WallPlan, or Graph2Plan would help clarify whether the observed performance gap persists across datasets. Also, how much does the additional preprocessing and postprocessing affect GSDiff’s results on MagicPlan?

  2. How are geometrically invalid predictions handled, and how frequently do they occur? As I mentioned in the last sectoin, are these outputs repaired, discarded, or included in evaluation? Could you report geometric validity statistics for FlowPlan, GSDiff, or MHD?

  3. Why does DFM outperform OE-DFM on all four metrics in Table 2? What explains the different behavior of these variants under footprint only generation and inpainting based conditional generation?

3 comments

We thank the reviewer for their constructive feedback. Before we proceed to address the questions we would like to report new improved results, achieved by performing stochastic inference using the same trained model, i.e. this is an inference time adjustment. We observe that injecting noise during the inference of continuous quantities (coordinates) substantially improves both quantitative and qualitative results. Our intuition is that this noise injection helps the model recover from sub-optimal trajectories in under-explored regions. Specifically, instead of performing a standard Euler step:

xtn+1=xtn+vθ∆tx_{t_{n+1}}= x_{t_n} + v_θ ∆t,

we adopt a two-step procedure, referring to this as Renoising:

  • we estimate the clean data sample: $ \hat x_1 = x_{t_n} + v_θ (1 −t_n) $
  • we perform a forward noising step at tn+1t_{n+1}: xtn+1=(1−tn+1)ε+tn+1x^1x_{t_{n+1}} = (1-t_{n+1})ε + t_{n+1}\hat x_1

Furthermore, regarding the discrete quantities, we found that applying corrector sampling[1], which increases the probability of re-masking a token at each step, also yields consistent improvements. Quantitative and qualitative ablation results on the MagicPlan dataset are reported in the table below and in Figure 12 in Appendix A.2

MagicPlan Dataset | Inference settings ablation

Model NameFID↓FID-Dinov2↓FID-Geom.↓IoU↑
OE-DFM6.7594.1126.860.94
OE-DFM w/ Renoising3.4131.4518.800.96
OE-DFM w/ Corrector Sampling6.6092.1826.440.94
OE-DFM w/ Renoising & Corrector Sampling3.0830.4918.300.96

[1]Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky T. Q. Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete flow matching. In Proceedings of the 38th International Conference on Neural Information Processing Systems, volume 37 of NIPS ’24, pp. 133345–133385, Red Hook, NY, USA, December 2024. Curran Associates Inc. ISBN 979-8-3313-1438-5.

Here, we are addressing the concerns of the reviewer.

Additional dataset and baseline

The reviewer requests further results on the RPLAN dataset and results from additional baselines. We provide new results comparing our method against FMLM and GSDiff on the MagicPlan and RPLAN datasets outperforming both approaches. On RPLAN, preprocessing and postprocessing steps for GSDiff are unnecessary because the ground truth samples do not contain wall gaps. Notably, the quantitative and qualitative results for GSDiff are very similar to those reported in the FMLM paper, despite being trained on different data splits. While the overall metrics suggest that RPLAN is a fundamentally simpler task, the performance gap between our approach and GSDiff remains.

MagicPlan Dataset

ModelType#Params (in millions)FIDFID-GeomFID-DinoIoU
GSDiffParallel13712.4927.82138.200.88
FMLMAutoRegressive1354.3114.6744.680.90
OurParallel1443.0818.3030.480.96

RPLAN Dataset

ModelType#Params (in millions)FIDFID-GeomFID-DinoIoU
GSDiffParallel13712.0510.6092.250.95
FMLMAutoRegressive1351.2411.5742.680.96
OursParallel1441.375.7116.520.99

Post processing of predicted polygons

The reviewer requests further clarification on the handling of predicted polygons.To address each of these concerns, we have included these details in the B and C section of the appendix. Specifically:

  • Invalid Polygons: While two-vertex rooms are discarded because they cannot form a valid polygon, the model very rarely generates such artifacts. We calculated the percentage of generated floorplans containing at least one discarded polygon.This occurred in only 0.2% of the MagicPlan dataset and 0.005% of the RPLAN dataset.
  • Boundary Violations & Uncovered Regions: Both issues are inherently penalized by the Intersection over Union (IoU) metric, as boundary violations increase the union while uncovered regions decrease the intersection. Furthermore, the MagicPlan ground truth data contains gaps (besides standard wall gaps) that may represent structural elements.
  • Self-Intersections: Self-intersections are not corrected during post-processing and appear as twisted polygons. These topological artifacts are penalized by the FID metric.
  • Room Overlap: We introduce a custom metric that calculates overlap within a floorplan as: Overlap=∑iArea(Pi)−Area(∪iPi)∑iArea(Pi)\text{Overlap} = \frac{\sum_{i}Area(P_i) - Area(\cup_iP_i)}{\sum_{i}Area(P_i)}, where PiP_i is the i-th generated polygon. The range of metric is [0,1][0,1] and lower values indicate less overlap. Our model achieves scores of 0.002 for MagicPlan and 0.003 for RPLAN, compared to 0.02 and 0.01 for FMLM, respectively. Given that the MagicPlan ground truth inherently contains a slight overlap of 0.0006, these results demonstrate that room overlap is negligible in the floorplans generated by both models. Finally, for GSDiff, the overlap is strictly 0.0 by construction due to its shared-nodes approach.

Here, we are addressing the remaining concerns.

Performance comparison between MHD and DFM on graph-conditioned generation

The reviewer addresses the gap between MHD and DFM on the graph-conditioned generation task and asks if training with custom masking can help performance. DFM employs the same architecture as MHD and has access to the same information(during inpainting) as MHD w/o Adj. The difference is that during training MHD w/o Adj always receives unmasked semantics while DFM receives only corrupted semantics. If we choose to train with partial masking (i.e. masking only the coordinates while keeping unmasked the semantics) essentially we replicate the training of MHD w/o Adj, with the difference that MHD is a diffusion-based model. In our new results we show superior performance over both MHD and MHD w/o Adj.

Performance comparison between OE-DFM and DFM on graph-conditioned generation

The reviewer addresses the gap between our models, OE-DFM & DFM, on the graph-conditioned generation task. The performance gap between OE-DFM and DFM is challenging to address. The difference between these two variants is that DFM reasons about coordinates of a specific room through positional embeddings(i.e. the order defines the connectivity) while OE-DFM by predicting corner indices. The performance gap could be the result of this modelling difference.

Difference of performance of DFM on footprint-conditioned & footprint-and-graph-conditioned generation

The reviewer indicates the difference in performance of DFM across footprint-only and inpainting-based conditional generation tasks and asks for a possible explanation. As described in section A.1 of the appendix, the discrepancy between these two tasks suggests that DFM cannot generate plausible semantics by its own, thus the unconditional generation, that depends on these, fails to generate plausible coordinates.

Reviewer_TCT8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FlowPlan, a unified continuous-discrete flow matching framework for polygonal floorplan generation. Conditioned on building footprints, the model jointly generates room geometry, structure, and semantics. The paper investigates different flow matching formulations and corner representations, including a permutation-equivariant design. Experiments on the MagicPlan dataset compare FlowPlan against two baselines, GSDiff and MHD, under different conditioning settings. The results demonstrate improved generation quality and inference efficiency compared to GSDiff, as well as competitive performance under additional test-time conditioning.

Critical strengths

This paper proposes a new formulation that unifies the generation of all attributes within a single framework. The design for handling both discrete and continuous attributes is reasonable, and the results in Table 1 demonstrate the potential of jointly generating these attributes.

Critical weaknesses

(1) The general idea of generating room geometry and semantics within a unified model has been explored in prior work, such as FMLM and CE2EPlan. Although FlowPlan adopts a different polygon representation and flow matching formulation, its distinction from these existing approaches could be better clarified.

(2) The contribution of adopting flow matching is limited, as such frameworks have been widely adopted in prior work. The extensive explanation and analysis of the architectural design seem unnecessary, since the distinctions between the discrete and continuous variants are straightforward. Instead, the experiments are insufficient to substantiate the claimed advantages. The comparison with GSDiff relies on additional heuristic pre- and post-processing, whose implementation details and potential effects on the reported performance are not sufficiently discussed. The comparison with MHD only demonstrates the applicability of the proposed approach rather than its effectiveness. Additional baselines, such as FMLM and Cons2Plan, should be included to further validate the advantages of the flow matching framework.

Questions for authors

(1) Why is it necessary to adopt flow matching, and what specific advantages does it offer? Additional baselines should be included to demonstrate its effectiveness.

(2) How was the model scaled from 33M to 144M parameters? Were the Transformer depth, hidden dimensions, number of attention heads, or other architectural components changed?

2 comments

We thank the reviewer for their constructive feedback. Before we proceed to address the questions we would like to report new improved results, achieved by performing stochastic inference using the same trained model, i.e. this is an inference time adjustment. We observe that injecting noise during the inference of continuous quantities (coordinates) substantially improves both quantitative and qualitative results. Our intuition is that this noise injection helps the model recover from sub-optimal trajectories in under-explored regions. Specifically, instead of performing a standard Euler step:

xtn+1=xtn+vθ∆tx_{t_{n+1}}= x_{t_n} + v_θ ∆t,

we adopt a two-step procedure, referring to this as Renoising:

  • we estimate the clean data sample: $ \hat x_1 = x_{t_n} + v_θ (1 −t_n) $
  • we perform a forward noising step at tn+1t_{n+1}: xtn+1=(1−tn+1)ε+tn+1x^1x_{t_{n+1}} = (1-t_{n+1})ε + t_{n+1}\hat x_1

Furthermore, regarding the discrete quantities, we found that applying corrector sampling[1], which increases the probability of re-masking a token at each step, also yields consistent improvements. Quantitative and qualitative ablation results on the MagicPlan dataset are reported in the table below and in Figure 12 in Appendix A.2

MagicPlan Dataset | Inference settings ablation

Model NameFID↓FID-Dinov2↓FID-Geom.↓IoU↑
OE-DFM6.7594.1126.860.94
OE-DFM w/ Renoising3.4131.4518.800.96
OE-DFM w/ Corrector Sampling6.6092.1826.440.94
OE-DFM w/ Renoising & Corrector Sampling3.0830.4918.300.96

[1]Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky T. Q. Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete flow matching. In Proceedings of the 38th International Conference on Neural Information Processing Systems, volume 37 of NIPS ’24, pp. 133345–133385, Red Hook, NY, USA, December 2024. Curran Associates Inc. ISBN 979-8-3313-1438-5.

Here, we are addressing the concerns of the reviewer.

Comparison with FMLM, CE2EPlan and Cons2Plan

The reviewer asks how our approach differs from FMLM and CE2EPlan. From existing prior work, FMLM and GSDiff represent the works most closely related to ours. However, a notable limitation of FMLM is its reliance on the coordinate grid resolution, which risks the loss of fine-grained geometric information. Mitigating this by increasing the coordinate grid resolution by just one decimal point drastically inflates the model size from 135M to 3.4B parameters.

Regarding CE2EPlan, it addresses a fundamentally simpler task: generating bounding boxes. Their final floorplans rely on a heavy post-processing step(snapping rooms to borders & neighboring rooms, overlaying bounding boxes to create more complex room layouts).

The reviewer mentions also Cons2Plan, which does not generate geometry and semantics concurrently. It requires semantics, in the form of a graph, as a conditioned input to the diffusion model. If semantic inputs are unavailable, they must be predicted from the boundary in an intermediate step, making it a multi-stage pipeline rather than a unified generative process.

Extensive architectural details

The reviewer highlights extensive details on the architectural details. We acknowledge that the sections detailing the architecture and modeling are extensive. The primary challenge of this work was formulating the floorplan generation problem to accommodate a variable number of rooms, each with a variable number of corners, which required a precise explanation. Furthermore, the underlying mechanics of Discrete Flow Matching (DFM) differ significantly from Continuous Flow Matching (CFM). DFM relies on Continuous-Time Markov Chains rather than ODEs, and predicts a rate matrix / posterior instead of velocities. Because combining these approaches is a novel application for this task, we felt a thorough theoretical grounding was necessary. However, to improve readability and conciseness, we have now consolidated the CFM and DFM paragraphs.

Insufficient experiments

Further, the reviewer mentions insufficient experimental results to back up our claims that we improve over baselines. We have included new results comparing our method against FMLM and GSDiff on both the MagicPlan and RPLAN datasets (Tables below). Notably, because GSDiff can be applied out-of-the-box on RPLAN, this comparison demonstrates that the observed performance gap is intrinsic to the proposed method itself, rather than an artifact of the preprocessing or postprocessing pipelines.

MagicPlan Dataset

ModelType#Params (in millions)FIDFID-GeomFID-DinoIoU
GSDiffParallel13712.4927.82138.200.88
FMLMAutoRegressive1354.3114.6744.680.90
OurParallel1443.0818.3030.480.96

RPLAN Dataset

ModelType#Params (in millions)FIDFID-GeomFID-DinoIoU
GSDiffParallel13712.0510.6092.250.95
FMLMAutoRegressive1351.2411.5742.680.96
OursParallel1441.375.7116.520.99

Comparison to Diffusion-based and AR-based baselines

The reviewer questions the necessity of the flow matching method for this task. Compared to diffusion-based approaches, flow matching offers improved training stability and faster inference, as the number of sampling steps is not predefined. Furthermore, unlike autoregressive (AR) models, parallel models decouple the number of inference steps from the length of the generated sequence. AR models also inherently require a specific room ordering, which is ill-defined for floorplans. Although FMLM mitigates this issue by employing permutation augmentation during training, parallel models can naturally circumvent the ordering requirement entirely as shown in our approach.

Clarifying scaling approach

The reviewer requests clarifying details about how we scaled our model. Since the approach is transformer-based, we scale the model by doubling the number of Transformer layers from 6 to 12 and increasing the model dimension from 512 to 768 without other modification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.