Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

FloorplanMAE:A self-supervised framework for complete floorplan generation from partial inputs

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FloorplanMAE claims a masked autoencoder with a lightweight Vision Transformer can complete residential floorplans from partial sketches, reporting better reconstruction metrics than pix2pix and cycleGAN.

desk verdict A plausible MAE-for-floorplans paper that is currently unverifiable because the quantitative table it depends on is missing; worth a re-submission, not referee time yet. read the letter →

arxiv 2506.08363 v2 pith:3E2DVLPA submitted 2025-06-10 cs.AI

Jun Yin , Jing Zhong , Pengyu Zeng , Peilin Li , Miao Zhang , Ran Luo , Shuai Lu This is my paper
classification cs.AI
keywords floorplangenerationmaskedautoencoderself-supervisedlearningVisionTransformerreconstructionpartialinputcompletionarchitecturaldesign
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that floorplan completion can be solved as a self-supervised image-reconstruction problem. It introduces FloorplanMAE, a masked autoencoder that hides patches of a floorplan and trains a lightweight Vision Transformer to predict what is missing, and FloorplanNet, a dataset of 7,700 standardized plus 300 real residential floorplans. If the central claim holds, architects working from partial sketches could get complete draft layouts automatically, and the same masked-reconstruction approach could scale to other kinds of drawings. The paper reports that FloorplanMAE outperforms pix2pix and cycleGAN on the reconstruction metrics it uses, and that real early-stage sketches are valid inputs.

What carries the argument

The load-bearing machinery is a masked autoencoder: the model hides a portion of a floorplan image, and a lightweight Vision Transformer encoder processes only the visible patches while an asymmetric decoder, armed with shared learnable mask tokens and positional embeddings, predicts the hidden patches. The training loss is the mean squared error between predicted and original masked patches. Masking and sampling strategy are part of the machinery: random, center, edge, one-sided, and corner masks simulate different partial-input scenarios, and the paper finds that higher masking ratios and dispersed masks change how much structure the model can recover.

What would settle it

An experiment that would settle the claim: train the identical FloorplanMAE model on floorplan images whose room labels and colors are randomly shuffled while wall thickness and color statistics are preserved; if reconstruction error barely changes, the model relies on texture statistics rather than architectural logic.

Watch

Extended reading notes

Core claim

The paper's central claim is that complete residential floorplans can be generated from partial inputs by treating the task as masked autoencoding. A lightweight Vision Transformer encodes only the visible patches of a floorplan image, an asymmetric decoder predicts the masked patches, and a mean-squared-error loss over masked positions drives training. The authors report that FloorplanMAE restores overall functional zoning and wall structure even at 80% random masking, with larger deviations in doors, windows, and exact room dimensions, and that it outperforms pix2pix and cycleGAN on the reconstruction metrics they report. They attribute the result to the model learning architectural design logic and functional zoning, not just local texture statistics.

Load-bearing premise

The method works only if masked autoencoding on rendered floorplan images teaches real architectural layout logic, not just filling blank patches by copying local visual patterns.

Editorial extensions

If this is right

  • Architects can feed an early partial sketch into FloorplanMAE and receive a complete residential floorplan draft, cutting repetitive redrawing work.
  • High masking ratios of 75 to 80 percent help the model grasp global structure, which is the opposite of the low masking ratios typical for natural images.
  • Mask placement matters more than masked content: contiguous masked regions produce blurrier reconstructions, while dispersed masks let the model infer missing areas more easily.
  • Real early-stage sketches can be used as inputs without relabeling, because pretraining is self-supervised on floorplan images.
  • FloorplanNet, with colored and line-drawing plans, provides a reusable benchmark for evaluating floorplan completion models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the paper leaves implicit is ablating the synthetic portion of FloorplanNet: performance on real hand-drawn sketches would show how much of the learned layout logic generalizes beyond the standardized synthetic generation tool.
  • Because the paper raises the question itself, a natural next step is measuring whether reconstruction metrics track functional zoning: two outputs with the same pixel error can have very different room circulations.
  • The same masking objective could likely transfer to other structured drawings such as facade sections, furniture layouts, or mechanical plans, but FloorplanMAE reports experiments only on residential floorplans, so that transfer remains speculative.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes FloorplanMAE, a masked-autoencoder framework with a Vision Transformer encoder and asymmetric decoder that reconstructs complete floorplans from masked raster inputs. It introduces FloorplanNet, a dataset of 8,000 residential floorplans (300 real, 7,700 generated by the RPLAN toolbox), and evaluates random, center, perimeter, one-sided, and corner masking strategies. The claimed result is that FloorplanMAE outperforms pix2pix and cycleGAN on 'various metrics' and produces plausible completions, including from real architectural sketches. The core mechanism is a standard MAE with a masked pixel MSE loss.

Significance. If the quantitative claim were substantiated, the contribution would be a straightforward but potentially useful application of masked autoencoding to floorplan reconstruction, with a modest dataset and a clear task framing. The paper does not claim architectural novelty beyond applying MAE, so its comparative value is as a baseline. The dataset and masking-strategy analysis could be useful to the community, but no code or dataset release is specified. As submitted, the central performance claim is not verifiable because the only quantitative evidence is a missing table, and pixel-level reconstruction is not tied to architectural structure. The paper also raises an interesting open question about FID and design rationality, but it is not developed.

major comments (3)
  1. [§4.4] The central claim that FloorplanMAE outperforms pix2pix and cycleGAN 'on various metrics' is not verifiable: the text states 'it can be seen from the table 1,' but no Table 1 is present in the manuscript. No metric names, numerical values, error bars, masking ratios, or baseline configurations are reported anywhere else. The authors should add the table, define each metric, report mean and standard deviation over the 500 test plans, and describe the baseline training settings.
  2. [§3, §4.3–4.4] The only training objective is masked pixel MSE, and all qualitative success is based on visual inspection. Because raster floorplans have strong local color and edge statistics, low MSE and plausible-looking completions do not demonstrate that the model has learned room adjacency, wall connectivity, or functional zoning, as claimed in the Introduction. The authors should add structural evaluation metrics (e.g., room adjacency graph correctness, wall continuity, functional-zone overlap) and/or a probe distinguishing reconstruction from interpolation, and should report a quantitative evaluation of the real-sketch validation, which is currently only qualitative.
  3. [§4.1] The dataset description is incomplete: the relationship between the 300 real and 7,700 RPLAN-generated plans and the 7,000/500/500 train/validation/test split is not stated, and the RPLAN toolbox is mentioned without a citation or license/access information. If the test set contains plans from the same synthetic toolbox as training, the generalization claim for real architectural sketches is weakened; the authors should report the composition of each split and evaluate on held-out real plans.
minor comments (5)
  1. [§4.3.5] The corner masking paragraph says it 'involves masking only one side of the floor plan,' which is inconsistent with corner masking; this should be corrected.
  2. [§3] The loss definition uses xi and x̂i for masked 'data points' but does not specify whether these are pixel values, patches, or normalized features; clarify the notation and the handling of masked versus unmasked pixels.
  3. [Global] The paper contains numerous typographical and formatting issues (e.g., 'desig n process' in the Abstract, 'asillustrated' in Section 4.3.4, and the stray white-box instruction on page 1), which should be corrected.
  4. [References] The reference list includes entries that are never cited in the body (e.g., FloorPlan-DeepSeek, UrbanSense, SAAF, ArchiLense); either cite them or remove them.
  5. [Figure 3] Figure 3 is labeled 'Example Results_Comparison of Different Masking Strategies' but is referenced as a general masking illustration; the caption should be aligned with what is actually shown.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the claim is an empirical benchmark comparison, and the missing Table 1 is an evidence gap rather than a self-referential reduction.

full rationale

The paper's central claim is an empirical reconstruction-performance comparison, not a derived equation. The only formal object is the masked-MSE loss in Section 3, L_MAE, and the pretraining/evaluation loop is standard: train on FloorplanNet's training split and report behavior on a held-out test split. No parameter is fitted to the reported outcome, no result is proved from a self-cited theorem, and the named baselines (pix2pix, cycleGAN) are external. The many self-citations (e.g., Yin et al. 2024, Zeng et al. 2024) are contextual or dataset-building references and are not load-bearing for the performance claim. I therefore find no circular step, and no equation or fitted value reduces by construction to the paper's own inputs. Two non-circular weaknesses are noted and should be weighed separately. Section 4.4 says 'it can be seen from the table 1 that our method outperforms them on various metrics,' but Table 1 is absent from the manuscript, so the quantitative superiority claim cannot be checked from the supplied text. In addition, the Introduction's assertion that the model must 'capture geometric structures and understand design logic and functional zoning' goes beyond what the pixel-MSE objective alone demonstrates. These are evidence and completeness problems, not instances of a claim reducing to its own inputs.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central contribution rests on three assumptions about the dataset, the representation, and the learning signal. No free parameters are fitted to the target result beyond standard training; the listed choices are hyperparameters and mask ratios selected by hand. No new physical entity is proposed. The FloorplanNet dataset is a new artifact whose independence cannot be checked because it is not released.

free parameters (3)
  • Masking ratio per strategy = random 80%, center 30%, perimeter 70%, one-sided 30%, corner 75%
    Hand-selected in Sections 4.2 and 4.3; the paper states higher masking ratios improve structural understanding for floorplans, but no search or sensitivity analysis is shown.
  • ViT architecture hyperparameters = not reported
    The encoder is described only as 'lightweight ViT'; patch size, depth, width, decoder depth, and mask token dimension are not given, yet they determine reconstruction quality.
  • Training hyperparameters = not reported
    Optimizer, learning rate, batch size, epochs, and data augmentation are absent, preventing replication.
assumptions (3)
  • domain assumption The FloorplanNet dataset is representative enough to validate floorplan completion for architectural practice
    Section 4.1 combines 7,700 synthetic RPLAN plans with 300 real plans, but does not characterize the test distribution or real-world diversity.
  • domain assumption Masked reconstruction on raster images teaches functional design logic, not only texture statistics
    Introduction and Section 3 assert the model understands zoning and structure; this is assumed rather than demonstrated.
  • domain assumption Image patches are an adequate representation for architectural constraints
    The method treats floorplans as grids of pixels in Section 3, implicitly assuming vector-level wall and door relationships survive rasterization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FloorplanMAE:A self-supervised framework for complete floorplan generation from partial inputs." pith.science (2026). https://pith.science/paper/3E2DVLPA

@misc{pith2026250608363,
  author       = {Pith},
  title        = {Pith review of: FloorplanMAE:A self-supervised framework for complete floorplan generation from partial inputs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3E2DVLPA}},
  note         = {Machine review of arXiv:2506.08363}
}
read the original abstract

In the architectural design process, floorplan design is often a dynamic and iterative process. Architects progressively draw various parts of the floorplan according to their ideas and requirements, continuously adjusting and refining throughout the design process. Therefore, the ability to predict a complete floorplan from a partial one holds significant value in the design process. Such prediction can help architects quickly generate preliminary designs, improve design efficiency, and reduce the workload associated with repeated modifications. To address this need, we propose FloorplanMAE, a self-supervised learning framework for restoring incomplete floor plans into complete ones. First, we developed a floor plan reconstruction dataset, FloorplanNet, specifically trained on architectural floor plans. Secondly, we propose a floor plan reconstruction method based on Masked Autoencoders (MAE), which reconstructs missing parts by masking sections of the floor plan and training a lightweight Vision Transformer (ViT). We evaluated the reconstruction accuracy of FloorplanMAE and compared it with state-of-the-art benchmarks. Additionally, we validated the model using real sketches from the early stages of architectural design. Experimental results show that the FloorplanMAE model can generate high-quality complete floor plans from incomplete partial plans. This framework provides a scalable solution for floor plan generation, with broad application prospects.

Figures

Figures reproduced from arXiv: 2506.08363 by the authors.

Figure 1
Figure 1. FloorplanMAE Architecture Diagram. Next, we propose a floor plan reconstruction method based on mask autoencoder, as [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ArchiLense: A Framework for Quantitative Analysis of Architectural Styles Based on Vision Large Language Models

    cs.CV 2025-06 reject novelty 5.0 of 10

    A new dataset and VLM-based pipeline for automatically describing and classifying architectural styles, with claims of 84.5% matching accuracy and 92.4% expert consistency, but the 92.4% figure is absent from the full text.

Reference graph

Works this paper leans on

2 extracted references · 2 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Bao, H., Dong, L., Piao, S., & Wei, F. (2021). Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254. Devlin, J. (2018). Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. Hu, R., Huang, Z., Tang, Y., Van Kaick, O., Zhang, H., & Huang, H. (2020). Graph2plan: Learning ...

    arXiv 2021

  2. [834]

    FloorPlan-DeepSeek (FPDS): A multimodal approach to floorplan genera tion using vector-based next room prediction

    Lu, S., Yan, X., Xu, W., Chen, Y., & Liu, J. (2016, June). Improving auditorium designs with rapid feedback by integrating parametric models and acoustic simulation. In Building Simulation (Vol. 9, pp. 235-250). Tsinghua University Press. Nauata, N., Hosseini, S., Chang, K. H., Chu, H., Cheng, C. Y., & Furukawa, Y. (2021). House-gan++: Generative adversar...

    arXiv 2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.