acceptodds
Under review as a conference paper at ICLR 2027

STARPlan: Stable Topology-Aligned Representations for Autoregressive Vector Floorplan Generation

Abstract

Graph-constrained vectorized floorplan generation requires connecting unordered input graphs with ordered geometric sequences. We present STARPlan, an autoregressive framework that learns topology representations from complete input bubble graphs and reuses them throughout geometry generation. Through masked graph pretraining, a graph encoder learns node representations independently of geometry generation and is then frozen to provide stable structural conditions for an autoregressive Transformer. A graph-derived generation order establishes correspondence between graph nodes and generated rooms and doors. We find that a simple strategy of reusing the same topology representations as both complete-graph conditions and object-aligned features effectively bridges graph structure and autoregressive sequence modeling. The complete set of node representations supplies global structural context through cross-attention, while the representation of each corresponding node augments the coordinate embeddings of its object. Experiments on RPLAN demonstrate state-of-the-art performance in layout diversity and topology compatibility across mixed-training and unseen-room-count settings, together with improved visual and functional realism in a user study.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
vector floorplan generation, graph-constrained generation, autoregressive generation, graph representation learning, structured generation, transformers

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.17 over 6 reviews, above 49% of the others.
Reviewer_A7GxOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a graph-enhanced representation based on auto-regressive paradigm on vectorized floorplan generation task. It adopts the design of GraphMAE2 on bubble diagram inputs to achieve a more powerful graph encoding, and utilizes a functional order to process the native unordered floorplan sequence. The model is trained end-to-end based on the canonical auto-regressive transformer pipeline with the reconstruction loss and geometric loss. Extensive experiments are conducted on RPLAN benchmark dataset under multiple (mixed, separate) setting following the typical evaluation protocol.

Critical strengths
  1. Overall the paper is well written and easy to follow, on motivation introduction, task formulation as well as experimental designs.
  2. The graph-based representation is suitable for the conditional generation task based on bubble diagram, since the connectivity constraint is an important rubric for measuring the fidelity of the interior floorplans.
  3. The ablation studies are comprehensive and well-designed.
Critical weaknesses
  1. To me the most critical downside is on the novelty of the method. Overall the paper's main contribution is a discriminative graph-enhanced representation on the node of the bubble diagram, however the key design follows an existing work of GraphMAE2, which diminishes the originality of the method. The remaining pipeline overall follows existing works (Liu et.al 2022 and LayoutTransformer), which limits the technical soundness of the paper.

  2. Although the proposed functional ordering (FO) achieves notable improvement from the ablation study, I have not fully aware of the key methodological difference from the prefix ordering used in Liu et.al 2022, which also ensures a deterministic order for generation different types of the rooms.

  3. The paper uses the causal Transformer based on Llama architecture, which is a more advanced design compared with the most relevant work Liu et at 2022 and LayoutTransformer. Both works are based on the earlier version of GPT-series. Although leveraging better foundation models are encouraged and understandable, it would be better to see the improvement gain brought from the architecture upgrade from the ablation study.

Questions for authors

Does LayoutTransformer shown in the experimental study mean the floorplan generation paper of Liu et.al 2022 or LayoutTransformer: Layout Generation and Completion with Self-attention? I suggest to make it clear before making an abbreviation.

Reviewer_XRJmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors address the mismatch between an unordered graph input and an ordered geometric output sequence in floorplan generation. They propose an autoregressive model that derives a deterministic object order from the input graph (FO) and applies it to both the conditioning and the output sequence for stable generation. To further improve performance, graph node representations are pretrained on the complete input topology before geometry decoding, kept fixed, and reused as stable conditions throughout autoregressive generation (PGR, GAR).

Critical strengths
  • The paper identifies a key limitation of existing autoregressive floorplan generation methods, the mismatch between unordered graph inputs and ordered output sequences, and addresses it with a simple yet effective solution, achieving substantial performance improvements.
  • The paper proposes pretraining a GraphMAE2-based graph encoder on floorplan graphs to extract topology features effectively. This training strategy could potentially be applied to other graph-conditioned sequential generation tasks.
Critical weaknesses

Insufficient justification for the autoregressive formulation.

The core motivation, the mismatch between unordered graph inputs and ordered outputs, arises only because the paper adopts autoregressive modeling. Diffusion methods such as HouseDiffusion avoid this issue by denoising all objects jointly. We would appreciate it if the authors could first explain why autoregressive generation is preferable to diffusion for this task. Without this justification, resolving the ordering problem may appear to address a problem introduced by the modeling choice itself rather than a meaningful contribution.

Missing comparison with recent methods.

All baselines (LayoutTransformer 2021, HouseGAN++ 2021, HouseDiffusion 2023) are at least three years old. Recent methods cited in the related work (e.g., Hu et al., 2025; Shiohara & Yamasaki, 2026) are not compared. We understand that some rely on additional boundary conditions or lack public code. Even so, we would appreciate it if the authors could explicitly state the reasons for their exclusion, or compare with reported results where the settings match. Without this, the claim of “significant improvements over the state of the art” is not fully supported.

The introduction is overly condensed.

  • It would be helpful to include a brief motivation for why this research problem matters and where it can be applied in practice.
  • The descriptions of FO and GAR are very brief, which makes the contributions difficult to grasp from the introduction alone.
  • Several terms, such as “stable conditions” and “identity of the graph node,” are used without definition or explanation of why they are needed. In addition, Pretrained Graph Representation (PGR), a key component of the method, is not introduced by name in the introduction.
Questions for authors
  1. Since the number of objects, and thus the sequence length, is already determined by the input graph, could the authors clarify why the ⟨EOS⟩ token is necessary?

  2. In Figure 1, the generated output tokens do not appear to be fed back as inputs for the next step, as is standard in autoregressive decoding. It would be helpful to revise the figure to reflect this.

  3. In Figure 2, the step following “Append interior door(s) to π” is not clearly indicated. Could the authors clarify where the flow proceeds after this step?

  4. It is somewhat unexpected that JGM performs worse than PGR. Could the authors explain the reason in more detail and clarify whether the two variants use different initializations? If JGM is randomly initialized while PGR uses pretrained weights, the comparison may not be entirely fair.

  5. Recent studies have explored the use of boundary information for vector floorplan generation. Could the authors clarify the rationale for not incorporating boundary information into their approach? In particular, why was this setting chosen, and what advantages does it offer over boundary-conditioned approaches?

Reviewer_HnDuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an autoregressive method for generating vectorized floorplans from bubble diagrams. It converts the input diagram into a room-and-door graph, encodes it using a pretrained and frozen graph encoder, and derives a graph-based generation order. A Transformer then generates room and door bounding-box coordinates autoregressively. The decoder accesses the complete set of graph representations through cross-attention, while the representation of the corresponding graph node is added to each object’s coordinate-token embeddings.

Critical strengths

The generation framework is coherent and clearly motivated. The ablation studies provide useful evidence that graph-derived ordering, pretrained topology representations, and node-specific alignment contribute complementary benefits to layout quality and topology compatibility.

Critical weaknesses
  1. The quantitative comparisons are limited to LayoutTransformer, HouseGAN++, and HouseDiffusion, published in 2021–2023. More recent relevant methods, particularly FMLM and Cons2Plan, are discussed but not evaluated. Differences in task specifications or unavailable implementations may complicate comparison, but the paper does not explain these exclusions. Comparisons under matched conditions, where feasible, or explicit justification for the baseline selection would strengthen the empirical conclusions.
  2. Evidence of scalability to more complex layouts remains limited. The held-out room-count experiments and eight-room ablations are useful, but evaluation covers at most eight rooms, and topology compatibility deteriorates as room count increases. It remains unclear whether the proposed conditioning and alignment mechanisms remain effective for substantially larger graphs, longer output sequences, or layouts with more repeated room categories and more complex connectivity. Further evaluation or an analysis of performance and computational cost as graph size increases would clarify the method’s practical scope.
Questions for authors
  1. Could the authors explain why more recent baselines were excluded, including whether implementation availability or specific differences in conditioning and evaluation prevented comparison?
  2. Have the authors evaluated layouts beyond eight rooms? If not, could they provide a more detailed analysis of how connectivity accuracy and computational cost vary with graph size and complexity?
  3. Since Functional Order (FO) is manually defined, could the authors compare it with alternative ordering schemes, such as breadth-first, depth-first, or category-based ordering, to clarify which design choices contribute to its effectiveness?
Reviewer_faaFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STARPlan, an autoregressive framework for generating vectorized floorplans from input bubble diagrams. The central motivation is that bubble diagrams represent room connectivity as unordered graphs, whereas autoregressive floorplan generation requires an ordered sequence of geometric objects. This creates a challenge in preserving complete graph topology and maintaining correspondence between input graph nodes and generated geometric objects throughout decoding. The proposed framework consists of three main components:

  1. Pretrained Graph Representation (PGR): The authors construct an attributed graph containing room and door nodes and employ GraphMAE2 to learn node representations through masked graph pretraining. The pretrained graph encoder is subsequently frozen and provides topology information from the complete input graph.
  2. Functional Order (FO): A deterministic graph traversal algorithm establishes the generation order. It starts from the living room, traverses connected rooms according to semantic categories and topology, inserts interior doors after visiting their incident rooms, and places remaining doors and the front door at the end. This establishes correspondence between graph nodes and generated objects.
  3. Graph-Aligned Representation (GAR): The representation of the corresponding graph node is projected into the decoder's embedding space and added to all four coordinate-token embeddings of the associated object. Meanwhile, cross-attention allows the autoregressive Transformer to access representations of the entire input graph.

The model predicts quantized bounding-box coordinates autoregressively, using a reconstruction cross-entropy loss and a geometric loss based on differences between pairwise IoU matrices. Experiments are conducted on the RPLAN dataset, containing more than 60,000 valid layouts after preprocessing. The authors compare against LayoutTransformer, HouseGAN++, and HouseDiffusion under mixed-training and separate-training settings. The reported results indicate that STARPlan achieves improvements in visual realism, functional realism, diversity, and topology compatibility. In particular, under mixed training, it reduces FID from 12.0 to 9.0 and graph edit distance from 1.5 to 0.8 compared with HouseDiffusion. Ablation studies suggest that FO improves diversity, PGR substantially improves topology compatibility, and GAR provides additional improvements through node-specific alignment. The paper's central claim is that reusing complete-graph representations as both global topology conditions and object-specific identities provides a simple yet effective bridge between unordered graph structures and autoregressive geometric generation.

Critical strengths

1.This paper targets a meaningful problem: the mismatch between unordered graph inputs and ordered autoregressive generation for floorplan synthesis. The proposed separation of global graph conditioning (PGR) and node-level graph alignment (GAR) is conceptually clear and technically coherent. The design uses cross-attention for full-graph access during decoding and injects node embeddings via GAR, without dynamically updating graph representations.

2.The ablation study thoroughly validates FO, PGR and GAR, showing their complementary gains on diversity and topology consistency. Evaluations cover human realism, FID for diversity and GED for topology preservation. The paper further demonstrates strong generalization to unseen room counts (5–8 rooms). The authors properly acknowledge the gap between the simplified setup and real architectural design.

Critical weaknesses

1.Limited methodological novelty and under-justified design choices. The core graph representation is built on GraphMAE2, and the generation pipeline largely follows existing works (Liu et al., 2022; LayoutTransformer). The difference between the proposed functional ordering (FO) and the prefix ordering in Liu et al. (2022) is not clearly articulated, and the gain attributable to the Llama-style causal Transformer upgrade over the earlier GPT-series baselines is not isolated in ablation. Moreover, the paper adopts autoregressive modeling, which itself introduces the "unordered graph vs. ordered output" tension; the choice of autoregressive generation over joint diffusion-based generation (e.g., HouseDiffusion) needs explicit justification.

2.Incomplete comparisons and under-developed presentation. All baselines (LayoutTransformer 2021, HouseGAN++ 2021, HouseDiffusion 2023) are at least three years old; more recent methods cited in the related work are not compared, and the reasons for their exclusion are not stated. The introduction is overly condensed: core modules (FO, GAR, PGR) and key terms ("stable conditioning," "node identity") are not defined or motivated, making the contributions hard to grasp.

Questions for authors

Q1. What is the essential difference between FO and the prefix ordering used in Liu et al. (2022)? Can you provide an ablation that separately isolates the performance gains from the Llama backbone upgrade and from the FO ordering strategy?

Q2. Why is autoregressive generation preferred over diffusion for this task? And for the recent methods cited (e.g., Hu et al., 2025; Shiohara & Yamasaki, 2026), why are they excluded from comparison — and where settings match, can you compare against their reported results?

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

The paper tackles the task of generating concrete floor plans, represented as a series of bounding boxes denoting rooms and doors, conditioned on a bubble graph representation of the house, where nodes represent different rooms and edges represent the walkways between them. The authors propose an encoder-decoder architecture, where the encoder is trained using a self-supervised objective and subsequently frozen, while the decoder is a standard autoregressive transformer. The paper introduces a hand-designed algorithm for converting a bubble graph into a sequence of rooms and doors, referred to as Functional Order (FO), along with a manual feature injection mechanism in the first layer of the decoder, where the node features from the frozen encoder are added to the token representations of their corresponding bounding boxes, referred to as Graph-Aligned Representation (GAR). The authors compare their approach against several baselines on the RPLAN dataset and show promising results.

Critical strengths

The architecture proposed in the paper is fairly simple and straightforward, and the main finding that, in such data-limited settings where models are trained from scratch, an encoder pretrained using self-supervised objectives such as MAE performs better than training the encoder and decoder end-to-end is very interesting. The paper also shows strong experimental improvements over the baselines and an appreciate set of ablations.

Critical weaknesses

The paper's writing is quite sparse on details, with several parts that require further explanation or discussion left unaddressed, making it difficult to understand and follow. Here are a few examples:

  • Functional Order (FO), which is presented as one of the main contributions, is only explained through a flowchart in Figure 2. However, this figure leaves many details unclear. For example, in the "Select next room" step, what happens if there are multiple candidate rooms? Which one is selected? This would matter if the algorithm is supposed to be deterministic and produce a unique outcome for each graph, but is that actually necessary? Same questions for doors. What happens if a room has multiple doors? Does their ordering matter? More generally, is the exact ordering itself important? If the main purpose of FO is to establish a correspondence between the node representations produced by the encoder and the bounding box tokens used in GAR, the ordering seems largely irrelevant. If on the other hand, it’s supposed to shrink the numerous number of serialization permutations down to a small number to allow the decoder to train more effectively, is FO special somehow or would  any other BFS/DFS-like walk on the graph work the same? This needs further clarification.
  • As far as I understand, according to lines 267-269, in the user studies, volunteers rank four samples, one from each method, based on visual and functional realism. Scores of 2, 1, -1, and -2 are then assigned according to these rankings, and the results are averaged to obtain the values reported in Table 1. As a result, the scores across the four methods should sum to zero. However, the two realism columns in Table 1 do not sum to zero. Why is this the case? This needs clarification.
  • The authors mention the FMLM work by Shiohara et al. [1], which appears to be one of the most recent and relevant related works, yet do not compare against it experimentally. What is the reasoning behind this choice? The same question applies to the work by Liu et al. [2]. The paper refers to the unconditional version of this approach without the refinement network as "base" in Table 3, but does not compare against the full method in Table 1. [1] Shiohara, Kaede, and Toshihiko Yamasaki. "Unified Vector Floorplan Generation via Markup Representation." arXiv preprint arXiv:2604.04859 (2026). [2] Liu, Jiachen, et al. "End-to-end graph-constrained vectorized floorplan generation with panoptic refinement." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022.
Questions for authors

Have the authors tried initializing the encoder from the self-supervised pretrained weights, but allowing it to be finetuned during training? Does freezing the encoder outperform this approach?

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

In this work, the authors introduce a method to convert a graph-based representation of 2D floor layouts (nodes being rooms and doors, edges being connections between rooms & doors) into vectorized layout diagrams. The proposed method tackles the challenge by first using a graph-based masked-autoencoder to learn a latent representation of the input graph. An autoregressive transformer then learns to predict vectorized room plans conditioned on this latent representation. The authors enhance this method by further conditioning each vector element generation on the corresponding node features from the graph. On a large dataset of house plans, this method is shown to be superior to prior art across various metrics.

Critical strengths
  1. The proposed method shows a notable improvement over prior art.

  2. The paper provides interesting ablative analysis that are really informative about the different design decisions. Particularly, I really appreciate the ablation on different architectures for the task (PGP, JMR, PGR; Table 3 (b)).

  3. The proposed method is relatively straightforward.

4.The inclusion of a user study including architecture experts makes the presented results more reliable.

Critical weaknesses
  1. Qualitative results are not very convincing - Figure 4, row 3 (ours) - It seems that the door is shared between two rooms.

  2. Why should this problem be solved with bbox generation, rather than programmatically? Our goal is to go from a bubble diagram to a vectorized floor plan. One can imagine having useful abstractions that avoid issues such as overlap etc. and also make the floorplan easier to manipulate / edit (for instance what should happen if I make the kitchen longer by moving the wall connecting it to another room? With a box-set representation as this paper uses, the resultant layout will have overlaps. For instance:

[1] Unified Vector Floorplan Generation via Markup Representation, Kaede Shiohara et al., CVPR 2026

  1. This method presents three contributions - functional order (FO), Pretrained Graph Representation (PGR), and Graph Aligned Representations. While each of these improve the quantitative results, its unclear what kind of errors each change helps address.
Questions for authors

I would appreciate the author’s view on the raised weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.