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

Jing Zhong; Jun Yin; Miao Zhang; Peilin Li; Pengyu Zeng; Ran Luo; Shuai Lu

REVIEW 4 major objections 5 minor 2 cited by

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

T0 review · 4 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 FloorPlan-DeepSeek recasts floor plan generation as autoregressive 'next room prediction' over vectorized room sequences, reporting competitive quality against diffusion models and Tell2Design.

desk verdict Plausible next-room prediction idea, but the missing evaluation section makes the central performance claim unverifiable. read the letter →

arxiv 2506.21562 v2 pith:RUVWHZWV submitted 2025-06-12 cs.CL cs.AIcs.AR

Jun Yin , Pengyu Zeng , Jing Zhong , Peilin Li , Miao Zhang , Ran Luo , Shuai Lu This is my paper · ORCID

classification cs.CLcs.AIcs.AR

keywords AutoregressivegenerationFloorplanNextroompredictionVectorrepresentationText-to-floorplanLargelanguagemodelsMultimodallearningEditablelayout

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 proposes that floor plan generation need not be an end-to-end, single-pass image synthesis task. It reframes the problem as 'next room prediction': given a text prompt and the rooms generated so far, a large language model predicts the next room as a vector of room type, center coordinates, width, and height. This turns a floor plan into a language-like sequence, so standard autoregressive training and decoding techniques apply directly. FPDS reports an FID of 16.465, PSNR of 76.564, and SSIM of 0.934 on text-to-floorplan benchmarks, positioning it competitively with diffusion models and Tell2Design. If the result holds, floor plan design tools could move from black-box one-shot generation to progressive, editable, interaction-friendly construction.

What carries the argument

The load-bearing object is the room vector $r_i = (t_i, x_i, y_i, w_i, h_i)$ together with the sequence factorization $P(\mathcal{S} \mid x, \theta) = \prod_{i=1}^N P(r_i \mid r_1, \dots, r_{i-1}, x, \theta)$. This factorization converts a spatial layout into a token stream that a language model can be trained on, making 'next room' prediction directly analogous to 'next token' prediction. The auxiliary constraint losses (functional adjacency, editability, outline boundary) are added as a weighted term $\lambda \mathcal{L}_{\text{cnst}}$ to the language modeling NLL loss, and inference uses top-K sampling or beam search over room sequences. The vector format is what carries the claim of editability: output is standardized geometric data rather than pixels.

What would settle it

Take the trained FPDS model, generate plans from a held-out set of text prompts, and measure the fraction of plans with rooms overlapping beyond a small tolerance and the fraction with rooms crossing the building outline. Then retrain or re-evaluate with the auxiliary constraint loss set to zero ($\lambda = 0$). If the violation rates do not rise, the central claim that FPDS's constraints ensure geometric validity fails.

Watch

Extended reading notes

Core claim

FloorPlan-DeepSeek's core claim is that a complete vector floor plan can be written as a sequence of room vectors $\mathcal{S} = [r_1, \dots, r_N]$ with $r_i = (t_i, x_i, y_i, w_i, h_i)$, and that generating a plan is the same kind of operation as generating a sentence. The model factorizes the probability of the plan as $P(\mathcal{S} \mid x, \theta) = \prod_{i=1}^N P(r_i \mid r_1, \dots, r_{i-1}, x, \theta)$, so each room is predicted conditioned on the text prompt and all previously generated rooms. A pretrained large language model is fine-tuned with this objective, together with auxiliary losses for functional adjacency, editability, and outline containment. The paper reports that the resulting model produces structurally coherent plans for one- to three-bedroom units, with quantitative FID, PSNR, and SSIM scores competitive with diffusion baselines. The intended contribution is a paradigm shift in how floor plan generation is modeled, not just a new architecture.

Load-bearing premise

The claim stands on the premise that a five-number room vector joined with ordinary next-token training can teach a language model global geometric validity; the paper states no explicit geometric loss or post-check that would enforce non-overlap and outline containment.

Editorial extensions

If this is right

Floor plan generation becomes progressive: the partial plan is always available between room predictions, so a user can inspect and modify intermediate states before the plan is complete.
Outputs are vector-based room descriptions, so they can feed directly into CAD-style downstream workflows instead of requiring manual redrawing from raster images.
Standard LLM decoding tools—top-K sampling and beam search—carry over to layout generation, giving a natural control knob for quality versus diversity.
Competitive FID, PSNR, and SSIM against diffusion models indicates that autoregressive sequence modeling is a viable alternative to one-shot generation for structured spatial outputs.
The same next-room formulation could support interactive human-in-the-loop design by inserting constraints between steps, a direct corollary of the autoregressive setup.

Reading between the lines

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

Because the room vector is resolution-independent, FPDS-style models could be trained or fine-tuned on zoning, circulation, and other high-level constraints without pixel supervision; the paper does not test this.
The room token stream could be extended to doors, windows, and furniture tags, turning a floor plan into a full semantic scene graph generated one component at a time; this extension is natural but not in the paper.
A direct ablation of the auxiliary constraint term would clarify whether geometric plausibility comes from the learned language-model prior or from the explicit losses; the paper's qualitative description leaves this attribution open.

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.

Desk Editor's Note

read the letter

Short version: the idea is worth ten minutes of thought, but the paper as written cannot support the claim that FPDS is competitive. The empirical section is one paragraph of three numbers with no baselines, no dataset, no protocol; and PSNR 76.5 is a red flag that the evaluation is likely degenerate. I wouldn't ask a referee to review it until the authors supply what's missing.

What's actually new: framing floor plan generation as 'next room prediction' with vectorized rooms (type, center, size) fed to a fine-tuned LLM. Tell2Design already used similar vector rooms for text-guided generation, but FPDS makes the autoregressive next-room prediction the explicit paradigm and names DeepSeek as the backbone. The three claimed advantages—progressive construction, interactive editing, vector output—are real if the method works, and the paper describes the training objective (conditional NLL plus auxiliary constraints) clearly enough to reproduce the idea, though not the details.

Soft spots, in order of severity.

First, the evaluation. The experiments section reports FID 16.465, PSNR 76.564, SSIM 0.934 and says 'across multiple datasets,' but gives no dataset names, no train/validation split, no sample count, no rasterization resolution, no metric implementation, no error bars, and no baseline numbers. For a claim of competitiveness against HouseDiffusion and Tell2Design, that is a showstopper.

Second, PSNR 76.564 is not believable for a generative model output vs a reference layout. Anything above about 60 dB means per-pixel errors below 1e-4 of full scale, typical of near-identical images, unlikely for novel generated floor plans. It suggests the metric was computed on something degenerate—perhaps rasterizing final result against the conditioning mask, or using tiny images. The paper gives no way to rule this out.

Third, the auxiliary constraint loss L_cnst is never defined. Eq (4) says total loss is L_nll + lambda L_cnst, but the three constraints are only named. That is a missing piece of the method.

Fourth, inference hyperparameters—top-K, beam width, temperature, lambda—are absent, and the qualitative Figure 4 has no input prompts or quantitative backing.

Minor points: the references contain duplicated entries, some names in the Author Contributions (H.Z., T.Z., J.L.) aren't in the author list, and the Related Work section leans on many self-citations, which is not a problem in itself but adds noise.

Verdict: the conceptual direction is reasonable and clearly explained, and the Discussion is honest about limitations. But the central empirical claim is unverifiable as written. I'd ask the authors to come back with a complete experimental section: datasets, baselines, hyperparameters, error bars, and a definition of L_cnst. Until then, it's not a paper I'd send to referees.

Referee Report

4 major / 5 minor

Summary. The paper proposes FloorPlan-DeepSeek (FPDS), an autoregressive 'next room prediction' framework for vector-based floorplan generation. Each room is encoded as a vector (type, center coordinates, width, height), and a pretrained large language model is fine-tuned to predict the next room conditioned on a textual prompt and previously generated rooms. The authors claim that FPDS achieves competitive performance against diffusion models and Tell2Design, reporting FID 16.465, PSNR 76.564, and SSIM 0.934 across 'multiple datasets'. The paper also introduces three auxiliary constraints (functional adjacency, editability, outline boundary) and an inference strategy using top-K sampling and beam search.

Significance. If the central empirical claim were substantiated, the paper would make a useful conceptual contribution: reformulating floorplan generation as autoregressive next-room prediction, which could support progressive, interactive, and vector-editable design workflows. The idea of adapting LLM-style autoregression to structured vector layout generation is timely and potentially valuable for intelligent architectural design. However, the significance is entirely contingent on the empirical evidence, which as reported is not verifiable. The paper does not provide baseline metric tables, dataset specifications, evaluation protocols, or hyperparameter settings, and the reported PSNR value is outside the plausible range for generative image comparison, raising serious concerns about the validity of the evaluation. The paper therefore cannot currently support its claimed contribution.

major comments (4)

[Experiment] The central claim of 'competitive performance' is unsupported because the Experiment section reports only FPDS's aggregate metrics (FID 16.465, PSNR 76.564, SSIM 0.934) with no baseline table, no dataset names, no train/test split, no number of evaluation samples, no image resolution or rasterization procedure, and no error bars or variance estimates. Without a controlled comparison against diffusion models and Tell2Design on identical data and metric implementations, the claim of competitiveness cannot be evaluated. This is a load-bearing omission for the paper's main contribution.
[Experiment (PSNR value)] The reported PSNR of 76.564 dB is implausibly high for comparing generated floorplans against reference images. PSNR values above roughly 60 dB correspond to per-pixel normalized differences below about 1e-4 and are typically observed only when the compared images are nearly identical by construction (e.g., reconstruction of the conditioning input or near-duplicate evaluation). The paper provides no details on how PSNR was computed, at what resolution, or what reference images were used, so the reader cannot rule out a degenerate evaluation setup. This specific value undermines confidence in the entire quantitative evaluation.
[Methodology, Eq. (4)] The auxiliary constraint losses named in the Methodology (Functional Adjacency, Editability, Outline Boundary) are asserted to be part of the optimization objective in Eq. (4) through the term ℒ_cnst(θ), but no functional form, weighting, or optimization procedure is given. Since the paper argues that these constraints ensure geometric plausibility (e.g., preventing rooms from exceeding the outline), and since no other geometric validity mechanism is described, the central premise that FPDS produces valid layouts is unverified. The reader cannot tell whether the constraints are actually implemented or how λ is chosen.
[Inference strategy] The inference strategy section states that top-K sampling and beam search are used, but the specific values of K, beam width, temperature, and the balancing coefficient λ from Eq. (4) are not reported. These hyperparameters are necessary for reproducibility and directly affect the reported quality metrics. Their absence, combined with the missing fine-tuning details (e.g., base model variant, training data, compute), makes the experimental results impossible to reproduce or interpret.

minor comments (5)

[Keywords] There is a typographical error in the keywords line: '.Autoregressive Generation' has a stray period; the list also appears to be formatted inconsistently with the rest of the paper.
[References] The reference for Tell2Design is duplicated in the reference list (entries under Leng et al., 2023), and several references in the text (e.g., Gao et al., 2023; Huang et al., 2025; Li et al., 2019) are cited in the introduction but do not appear to correspond to the reference list entries in a way that is checkable from the manuscript text.
[Figure 4] Figure 4 is described as a qualitative comparison with Obj-GAN, Tell2Design, Glm-4-9b-chat, Qwen2.5-VL-72B-Instruct, and FPDS, but the figure itself is not visible in the submitted text and the input prompts are 'omitted due to length'. Without the figure and prompts, the qualitative claims cannot be assessed.
[Authorship information] The author contributions section lists individuals (H.Z., T.Z., J.L.) who are not listed as authors on the title page, creating an inconsistency in the manuscript provenance.
[Discussion] The Discussion section acknowledges that the model 'struggles with layout errors under complex semantics' and that 'local mismatches remain', which is in tension with the strong quantitative claims in the Experiment section; this tension is not reconciled anywhere in the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FPDS is a standard autoregressive sequence model applied to room vectors; its reported metrics are empirical measurements, not quantities forced by the model's definitions.

full rationale

The derivation chain is: rooms are encoded as vectors (Eq. 1), ordered into a sequence (Eq. 2), and modeled by the conditional autoregressive likelihood P(S|x,theta) = prod P(r_i | r_<i, x, theta) (Eq. 3), trained with an NLL objective plus an unspecified auxiliary constraint term (Eq. 4). This is the ordinary next-token-prediction objective used by LLMs, applied to room tokens; the paper does not claim to derive floorplan quality from this objective alone, and the objective does not contain the evaluation metrics (FID, PSNR, SSIM) as terms. The reported numbers are empirical outputs of a trained model, not constants folded into the equations. The paper's 'next room prediction' framing is explicitly presented as an adaptation of LLM autoregression, citing GPT and DeepSeek, so there is no hidden import of a conclusion. No load-bearing self-citation is used: the many works by the same authors appear in the introduction and reference list as general related work, and the central method cites Tell2Design (external) for the vectorized representation. The auxiliary constraints in the Methodology are described only qualitatively and Eq. 4 leaves L_cnst unspecified, which is an incompleteness/verifiability problem, not a circularity: nothing in the paper shows a fitted parameter being renamed a prediction or a target metric being used to define the model. The evaluation section lacks baselines and protocol details, but absent evidence that the metrics were computed from the training objective or from the conditioning input, this does not constitute demonstrated circularity. Overall the central empirical claim rests on training and evaluation of a model, not on a self-referential derivation.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical entities, forces, or conserved quantities are introduced; the contribution is a modeling paradigm rather than a new entity. The free parameters and axioms listed above are the elements the central claim depends on but does not provide enough information to verify.

free parameters (2)

lambda (constraint loss weight) = not reported
Appears in Eq. (4) as the balance between language modeling loss and auxiliary constraint losses; no value or tuning procedure is given, yet it controls how strongly geometric constraints influence training.
Top-K value and beam width = not reported
Inference strategies in the Inference strategy section are stated without concrete values, although they affect quality and diversity.

assumptions (5)

domain assumption A floor plan can be represented as an ordered sequence of room vectors (type, center, width, height) without loss of structural information.
Eqs. (1)-(2) define the representation; the paper assumes this sequence captures all information needed for layout generation, including adjacency and topology.
standard math Conditional factorization P(S|x,theta) = product_i P(r_i | r_<i, x, theta) is a valid generative model for floor plans.
This is the standard chain rule for autoregressive models; it is valid for any ordering, but the paper assumes a meaningful room order exists and can be learned.
domain assumption Pretrained DeepSeek-R1-32B, after fine-tuning, can output structured room vectors and capture spatial relationships through multi-head attention without explicit geometric inductive bias.
Methodology states fine-tuning a pretrained LLM; no architectural modifications or specialized geometric heads are described.
ad hoc to paper Auxiliary constraint losses can be expressed and optimized, even though their functional forms are not given.
The Auxiliary constraints section lists three qualitative constraints and Eq. (4) refers to a total constraint loss L_cnst, but the actual loss functions are never defined; the central claim depends on their existence and effectiveness.
domain assumption The reported FID, PSNR, and SSIM values were computed on a comparable evaluation protocol to the baselines.
The Experiment section gives no dataset splits, metric computation details, or baseline numbers, yet the claim of competitiveness relies on this comparability.

how reviews work

0 comments

Cite this review

Pith. "Pith review of FloorPlan-DeepSeek (FPDS): A multimodal approach to floorplan generation using vector-based next room prediction." pith.science (2026). https://pith.science/paper/RUVWHZWV

@misc{pith2026250621562,
  author       = {Pith},
  title        = {Pith review of: FloorPlan-DeepSeek (FPDS): A multimodal approach to floorplan generation using vector-based next room prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RUVWHZWV}},
  note         = {Machine review of arXiv:2506.21562}
}

read the original abstract

In the architectural design process, floor plan generation is inherently progressive and iterative. However, existing generative models for floor plans are predominantly end-to-end generation that produce an entire pixel-based layout in a single pass. This paradigm is often incompatible with the incremental workflows observed in real-world architectural practice. To address this issue, we draw inspiration from the autoregressive 'next token prediction' mechanism commonly used in large language models, and propose a novel 'next room prediction' paradigm tailored to architectural floor plan modeling. Experimental evaluation indicates that FPDS demonstrates competitive performance in comparison to diffusion models and Tell2Design in the text-to-floorplan task, indicating its potential applicability in supporting future intelligent architectural design.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

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.
FloorplanMAE:A self-supervised framework for complete floorplan generation from partial inputs
cs.AI 2025-06 conditional novelty 4.0 of 10

FloorplanMAE uses masked autoencoding to reconstruct complete residential floorplans from partially masked inputs, reporting qualitative gains on synthetic and real floorplans.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages · cited by 2 Pith papers

[1]

H., Cheng, C

Nauata, N., Chang, K. H., Cheng, C. Y., Mori, G., & Furukawa, Y. (2020). House-gan: Relational generative adversarial networks for graph- constrained house layout generation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16 (pp. 162-177). Springer International Publishing. https://doi.org/10.1...

arXiv doi:10.48550/arxiv.2311.15941 2020
[834]

UrbanSense:A Framework for Quantitative Analysis of Urban Streetscapes leveraging Vision Large Language Models

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. Yan, X., Lu, S., & Li, J. (2016). Experimental studies on the rain noise of lightweight roofs: Natural rains vs artifi...

work page Pith review arXiv 2016
[1674]

Street microclimate prediction based on Transformer model and street view image in high-density urban areas

https://doi.org/10.3390/buildings15101674 Li, J., Lu, S., Wang, W., Huang, J., Chen, X., & Wang, J. (2018). Design and Climate‐ Responsiveness Performance Evaluation of an Integrated Envelope for Modular Prefabricated Buildings. Advances in Materials Science and Engineering, 2018(1), 8082368. Li, J., Song, Y., Lv, S., & Wang, Q. (2015). Impact evaluation ...

work page doi:10.3390/buildings15101674 2018

Pith tools

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

[1] [1]

H., Cheng, C

Nauata, N., Chang, K. H., Cheng, C. Y., Mori, G., & Furukawa, Y. (2020). House-gan: Relational generative adversarial networks for graph- constrained house layout generation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16 (pp. 162-177). Springer International Publishing. https://doi.org/10.1...

arXiv doi:10.48550/arxiv.2311.15941 2020

[2] [834]

UrbanSense:A Framework for Quantitative Analysis of Urban Streetscapes leveraging Vision Large Language Models

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. Yan, X., Lu, S., & Li, J. (2016). Experimental studies on the rain noise of lightweight roofs: Natural rains vs artifi...

work page Pith review arXiv 2016

[3] [1674]

Street microclimate prediction based on Transformer model and street view image in high-density urban areas

https://doi.org/10.3390/buildings15101674 Li, J., Lu, S., Wang, W., Huang, J., Chen, X., & Wang, J. (2018). Design and Climate‐ Responsiveness Performance Evaluation of an Integrated Envelope for Modular Prefabricated Buildings. Advances in Materials Science and Engineering, 2018(1), 8082368. Li, J., Song, Y., Lv, S., & Wang, Q. (2015). Impact evaluation ...

work page doi:10.3390/buildings15101674 2018