{"id":"82ac03ca-0598-4852-8934-d61147f41844","arxiv_id":"2501.09279","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A Stable Diffusion model fine-tuned with LoRA and guided by ControlNet generates residential layouts from text prompts that describe room counts, sizes, and connections, with claimed improvements over two baselines.","lead":"This paper uses Stable Diffusion, LoRA, and ControlNet to generate residential floor plans from natural-language prompts plus either an example layout or a boundary image. It claims this multimodal approach is more flexible than existing layout generators and outperforms Pix2Pix and HouseDiffusion, but the reported evaluation has serious methodological gaps.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation protocol cannot support 'better than SOTA': Table 5 lacks a held-out split and a matched-pair protocol, so high PSNR/SSIM/LPIPS may measure memorization of RPLAN training images rather than constraint satisfaction.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: the evaluation in §5.2.1 does not establish that the model generalizes or that it outperforms state-of-the-art methods. My reading of the full text reinforces this. The quantitative comparison is the only direct support for the abstract's claim of superiority under multimodal constraints, and that comparison is uncontrolled in three separate ways: there is no described train/test separation, no pairing rule for the 8,000 \"pairs,\" and no evidence that the baselines were trained and evaluated under identical conditions. Each of these alone would be serious; together they make Table 5 uninterpretable as evidence of constraint satisfaction or of comparative advantage. The qualitative figures and ablation studies are suggestive but not convincing: the ablations explicitly document failures to control connections and area proportions, and Section 5.4 concedes the same limitation. I do not see an internal inconsistency in the method itself as the primary issue; rather, the paper's central empirical claim is not supported by the reported evidence. The one concrete check that would settle the concern is a properly split, matched-pair, comparably-baselined evaluation with constraint-accuracy metrics. If that check were run and confirmed the table's numbers, the central claim would be substantially rehabilitated; as reported, the claim remains unestablished.","tokens_in":20814,"tokens_out":3390,"duration_ms":61342,"concrete_test":"Split RPLAN into train/validation/test sets by plan identity; fine-tune LoRA and ControlNet on the training split only. For each held-out test prompt and boundary, generate multiple samples and compute: (i) FID between the full set of generated test images and the full set of held-out real test images, not paired images; (ii) PSNR, SSIM, and LPIPS only for samples conditioned on that specific held-out plan, matched to the corresponding ground truth; and (iii) constraint-satisfaction metrics: room-count error, connectivity graph precision/recall, and area-ratio error. Retrain Pix2Pix and HouseDiffusion on the same split and evaluate them under the same protocol. If Table 5's scores are reproduced on held-out data and baselines are comparably trained, the central claim survives; if the scores collapse or the baselines were not comparably trained, it does not.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Table 5, but the evaluation protocol is unspecified in a way that can invert the conclusion. Section 5.2.1 reports only that \"we randomly selected 8,000 pairs of generated and real images\"; it never states that the real images were held out from training, how pairs were matched, or whether the Pix2Pix and HouseDiffusion baselines were retrained on the same RPLAN split with the same text and boundary conditions. Since the RPLAN dataset is the training corpus, a LoRA model that merely memorizes training images would score near-perfectly on PSNR, SSIM, and LPIPS when compared against those same images as \"real\" references, while having zero flexibility. FID, a distribution-level metric, is also reported as if computed on the same 8,000 paired samples; the low FID values could reflect distribution overlap with training data rather than constraint satisfaction. The paper's own ablations (§5.3) show uncontrolled connectivity and generation failures, and §5.4 concedes that \"current knowledge representations are still insufficient in terms of area and connectivity control.\" With no train/test split, no pairing protocol, no baseline training details, and no released artifacts, the quantitative evidence does not establish the claimed superiority over state-of-the-art models.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a residential layout generation method based on a Stable Diffusion model fine-tuned with LoRA, combined with ControlNet for boundary conditioning and a knowledge-graph-to-natural-language scheme for encoding design rules. The method supports two generation pathways: one conditioned on text plus a reference floor-plan image, and one conditioned on text plus a boundary image processed by Canny edge detection. The authors evaluate their method on the RPLAN dataset against a non-fine-tuned SD1.5 baseline, Pix2Pix, and HouseDiffusion using FID, PSNR, SSIM, and LPIPS, and they present ablation experiments in which area, connectivity, or room-count information is omitted. The central claim is that the proposed method provides more flexible and controllable residential layout generation than state-of-the-art models, even when semantic information about room areas or connections is incomplete.","tokens_in":20918,"tokens_out":4729,"duration_ms":47943,"significance":"If the claimed results were supported, the paper would make a useful contribution to AI-based architectural layout design: it addresses a real limitation of prior work by accepting natural-language design constraints, it provides a concrete mapping from a knowledge graph to text prompts, and it demonstrates the feasibility of fine-tuning a large text-to-image model for a structured design task. The work also benefits from using an external public dataset (RPLAN) and from being transparent about failure modes, including the concession in §5.4 that area and connectivity control are still insufficient. The central quantitative claim, however, is not yet established because the evaluation protocol is underspecified and the reported metrics do not measure the claimed constraint-satisfaction and flexibility properties.","major_comments":[{"comment":"The evaluation protocol cannot support the central claim of superiority over state-of-the-art methods. The text states only that \"we randomly selected 8,000 pairs of generated and real images to evaluate specific metrics,\" but it does not state that the real images were held out from training, how pairs were matched, or whether the baselines were trained on the same data with the same prompts and boundary conditions. Because the RPLAN dataset is the training corpus, the high PSNR, SSIM, and LPIPS values in Table 5 could reflect memorization of training images rather than constraint satisfaction. FID is a distribution-level metric and is not defined for paired samples; a low FID could also reflect overlap with the training distribution. The abstract's claim that the method is \"better than state-of-the-art models\" is therefore not established by Table 5.","section":"§5.2.1, Table 5"},{"comment":"The comparison against Pix2Pix and HouseDiffusion is not reproducible because no details are given about how these baselines were configured, trained, or evaluated. It is not stated whether they were retrained on the same RPLAN preprocessing, whether they received the same text and boundary inputs, or whether the same held-out prompts were used. Table 5 also reports single mean values without error bars, confidence intervals, or significance tests, so the statements in §5.2.1 about \"minimal fluctuation and high stability\" are unsupported. Without a matched evaluation protocol, the numerical gaps in Table 5 cannot be interpreted as evidence of the proposed method's superiority.","section":"§5.2.2 and §5.1"},{"comment":"No quantitative metric measures the properties at the core of the paper's claims: room-count accuracy, connectivity correctness, boundary adherence, or constraint satisfaction. The ablation figures themselves show failures, including generation cases with incorrect room connectivity (§5.3.1), balconies connected to the kitchen instead of the living room (§5.3.2), and uncontrolled connections when connection information is missing. Section 5.4 explicitly concedes that \"current knowledge representations are still insufficient in terms of area and connectivity control.\" The qualitative figures cannot substitute for a constraint-satisfaction metric, and the claim that the method is flexible and controllable under incomplete semantic information is not quantitatively supported.","section":"§5.3 and §5.4"},{"comment":"The description of ControlNet in Eq. (4) as a dynamic per-layer weight update W' = W + α · ZeroConv(C(B)) does not match the standard ControlNet mechanism, in which a trainable copy of U-Net blocks is injected via zero convolutions and the original weights remain frozen. As written, the equation is not an implementable description of the method and prevents reproducibility. If the authors implemented a different mechanism, Eq. (4) and Fig. 7 must be corrected to state what was actually done; otherwise, the boundary-conditioning pathway is not reproducible.","section":"§4.2.2, Eq. (4)"}],"minor_comments":[{"comment":"The descriptions of Path 1 and Path 2 in §5.2.2 appear to reverse the definitions given in §4.2: in §4.2 the first path uses text plus a preferred floor-plan image, while the second path uses text plus ControlNet-processed boundaries, but the experimental text says the first path employs a control net and the second path leverages real images. Please reconcile the notation.","section":"§5.2.2"},{"comment":"The FID formula in Eq. (8) is not written correctly: the first term should be the squared Euclidean norm of the mean difference, and the covariance terms should use the standard FID expression. The variables μ, μw, Σ, and Σw are also not fully defined in the text.","section":"§5.1, Eq. (8)"},{"comment":"In Eq. (5), the roles of r and r' are unclear: if r' denotes the rank of the low-rank matrices, the scaling factor r/r' is not the standard LoRA scaling (which is typically α/r with respect to the rank), so the meaning of each symbol should be clarified.","section":"§4.2.3, Eq. (5)"},{"comment":"Fig. 10 is difficult to read in grayscale, and the individual metric distributions are not legible; reporting the standard deviations or interquartile ranges alongside Table 5 would be more informative.","section":"Fig. 10"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the general methodology is plausible, but the empirical evaluation as presented is not sufficient to support the abstract's claim of state-of-the-art performance. The main issues are correctable in principle: the authors would need to re-run the experiments with a clearly documented train/test split, a matched-pair protocol, properly computed distribution-level metrics, and quantitative constraint-satisfaction metrics, and also provide details on baseline training. Given that the load-bearing weakness is in the evaluation rather than in the underlying derivation, I view major revision as more appropriate than outright rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The reader's take is mostly right: the evaluation in Section 5.2.1 does not support the central claim. But I'd give the paper more credit than the reject verdict alone implies. The knowledge-graph-to-natural-language prompt scheme (Section 4.1.4) is a genuinely practical contribution. Translating room types, areas, and connectivity into structured text that a diffusion model can condition on is the kind of thing people in the AEC-text-to-floorplan space will want to build on. The two-path conditioning design (text+boundary vs. text+layout image) is also a sensible application of existing components, even if none of the components are new. The soft spots are as identified, and they are load-bearing. No train/test split is described; Section 5.2.1 only says 8,000 pairs were randomly selected. Since RPLAN is the training corpus, the PSNR/SSIM/LPIPS numbers could simply reflect memorization. FID is a distribution-level metric, so reporting it on paired samples is conceptually muddled. Table 5 has no error bars or significance tests, and the baselines are not described as being retrained on the same split. Omitting HouseGAN++ and Graph2Plan from the comparison is also a serious gap, since those are the closest graph-constrained baselines cited in the paper. The internal inconsistencies (Eq. 4's ControlNet formula, the reversed path descriptions in 5.2.2) suggest a hasty write-up. And no code or data is released, which makes the reproducibility situation worse. I don't see derivation-level circularity or invented entities. The paper's own ablations, though, are honest about failures: uncontrolled connectivity and area proportions under missing constraints, and Section 5.4 concedes the knowledge representation is insufficient for area and connectivity control. That honesty counts for something, but it also undercuts the abstract's claim about working well when semantic information is incomplete. Who is this for? Researchers working on text-conditioned floorplan generation, especially in the AEC community. They will find the prompt-design scheme useful and the baseline comparison instructive, but they should not trust the quantitative claims as reported. The paper deserves a serious referee rather than a desk reject because the core idea is worth fixing. I would recommend sending it to review with a request for major revision: a proper train/test split, matched-pair evaluation with significance tests, stronger baselines, corrected equations, and released artifacts. If those are addressed, this could become a citable application paper.","headline":"The prompt-template idea is genuinely useful, but the evaluation cannot support the 'better than SOTA' claim; with a corrected protocol and released artifacts this could be a solid application paper.","tokens_in":769,"tokens_out":889,"would_cite":false,"duration_ms":22550,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Text-conditioned, boundary-aware Stable Diffusion with LoRA fine-tuning generates residential layouts that satisfy multimodal constraints better than Pix2Pix and HouseDiffusion, even when room-area or connection details are missing.","keywords":["Residential Layout Generation","Multimodal Generative Design","Natural Language Processing","Stable Diffusion Model","Knowledge Graph","LoRA fine-tuning","ControlNet","Floor plan generation"],"falsifier":"Re-run the evaluation with the 8,000 comparison references drawn exclusively from a hold-out subset never used in LoRA fine-tuning, and report the four metrics separately for that unseen subset; if the LoRA-tuned model's advantage over HouseDiffusion and Pix2Pix shrinks substantially or reverses, the flexibility claim is falsified.","tokens_in":20473,"feed_emoji":"🏠","tokens_out":9223,"duration_ms":78171,"temperature":0.7,"pith_summary":"The paper proposes a flexible floor-plan generator that accepts natural-language descriptions (room counts, room types, sizes, adjacencies) together with either a reference layout image or a boundary drawing, and returns a residential layout that follows both. Its goal is to lower the design threshold so that non-experts can specify a home in plain sentences instead of drawing bubble diagrams or building knowledge graphs. To do this, the paper maps seven residential design rules into a knowledge graph and then into standardized text prompts, and fine-tunes a Stable Diffusion model with Low-Rank Adaptation (LoRA) while using ControlNet to enforce boundary conditions. The reported experiments compare the LoRA-tuned model against Pix2Pix and HouseDiffusion on FID, PSNR, SSIM, and LPIPS, and the ablation studies show that when area or connectivity information is omitted, the model produces diverse layouts rather than failing outright.","feed_headline":"Text-to-floor-plan model beats Pix2Pix and HouseDiffusion","feed_subtitle":"LoRA fine-tuning lets a Stable Diffusion model turn room descriptions and boundaries into residential layouts.","key_machinery":"The load-bearing machinery is the pairing of three components. (1) A knowledge-graph-to-text mapping converts each floor plan's room nodes and connectivity edges into a standardized sentence sequence (room types and counts, room areas, then adjacency relations), making design rules available as natural-language conditioning. (2) Low-Rank Adaptation (LoRA) fine-tunes the frozen Stable Diffusion U-Net and text encoder on recolored, upsampled RPLAN data, so the pretrained model learns floor-plan style and rule-following behavior with a small trainable parameter set. (3) ControlNet, conditioned on Canny-edge boundary images, adjusts the U-Net features to keep the generated layout inside the user-supplied boundary, while CLIP embeds the text prompt into the same latent space as the images. The two pathways differ only in the image condition: a full reference layout for Path 1 versus a boundary sketch for Path 2.","core_discovery":"The central claim is that a general-purpose text-to-image diffusion model can be specialized into a cross-modal residential layout generator that is more flexible than prior graph- or boundary-conditioned models. The paper implements this with two generation paths sharing the same text conditioning: Path 1 uses a natural-language prompt plus a reference floor plan image, and Path 2 uses the prompt plus a Canny-detected boundary image fed through ControlNet. Design knowledge is extracted from the RPLAN dataset into a knowledge graph whose nodes carry room type, position, and area and whose edges encode connectivity; this graph is then rendered as structured sentences such as 'bedroom_1 connect bathroom', which form the prompt. The paper reports that the LoRA-tuned model achieves FID of 22.1 (Path 1) and 28.5 (Path 2), versus 32.6 for HouseDiffusion and 142.6 for Pix2Pix, with higher PSNR and SSIM and lower LPIPS on both paths. The ablation experiments are used to argue that omitting room-area or room-connection information degrades strict compliance but leaves the model able to invent plausible layouts, which the paper presents as evidence of flexibility under incomplete semantic information.","pith_inferences":["Editorial extension: the reported metrics compare generated images to real RPLAN drawings at the pixel and feature level; they do not directly measure whether a generated plan contains exactly the requested number of rooms or the requested adjacencies, so a constraint-satisfaction metric would be a sharper test of the flexibility claim.","Editorial extension: the knowledge-graph-to-text recipe is specific to residential rules, but the same pipeline—extract a graph, render it as sentences, fine-tune a diffusion model with LoRA—could be ported to other layout problems such as office or hospital planning, provided the rule tables and vocabulary are rebuilt.","Editorial extension: the ablation figures show the model sometimes repairs missing constraints in plausible but unpredictable ways, such as connecting a balcony to the kitchen when the living-room connection was omitted; a deployed tool would likely pair the generator with a rule checker that flags such violations."],"forward_implications":["If the method works as reported, a homeowner can state 'three bedrooms, one balcony, living room connects kitchen' and receive a plausible floor plan without preparing bubble diagrams, room masks, or a knowledge graph.","The shared text conditioning means the same fine-tuned model serves both the reference-image path and the boundary-only path, so a designer can begin with a rough boundary and later supply a preferred precedent layout without retraining.","The ablation results imply the generator can operate with partial constraints: missing area descriptions or missing connectivity edges lead to diverse completions rather than a refusal, which is useful in early design stages when the program is not fully fixed.","The reported metric gains would move the practical baseline for floor-plan generation from GAN-based image translation and vector diffusion toward fine-tuned latent diffusion with natural-language control."],"supporting_citations":[{"why":"The diffusion-based state-of-the-art baseline whose FID, PSNR, SSIM, and LPIPS scores the proposed method claims to beat.","marker":"[7]"},{"why":"The image-to-image GAN baseline that the paper compares against, representing the GAN-based approach it claims to surpass.","marker":"[56]"},{"why":"Supplies the RPLAN dataset and the graph-to-floor-plan method that the proposed natural-language representation extends and contrasts with.","marker":"[38]"},{"why":"A graph-constrained layout generation method that represents the earlier single-input graph-based approach the paper critiques.","marker":"[5]"},{"why":"The graph-constrained GAN method with relational constraints, used as a prior-art anchor for design-rule representation.","marker":"[6]"},{"why":"A prior diffusion-based multi-conditional floor-plan generator that the paper positions against to highlight its flexibility in input types.","marker":"[55]"},{"why":"Prior work evaluating text-to-image tools for floor plans, which motivates using natural language as the design-constraint modality.","marker":"[26]"},{"why":"A graph-constrained high-rise residential floor-plan generation method from the same domain that the proposed approach extends.","marker":"[4]"}],"fun_headline_variants":["Stable Diffusion turns text and sketches into flexible floor plans","Text-to-layout model beats Pix2Pix and HouseDiffusion with LoRA","Cross-modal diffusion flexibly designs rooms from language and edges","Flexible residential layouts from text prompts via Stable Diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation's validity rests on the assumption that the 8,000 real floor plans used as references were not seen during training and that looking like those drawings is a good proxy for design quality; the paper states only that it randomly selected 8,000 pairs, with no train/test split or pairing protocol.","fun_headline_variants_meta":{"raw":{"variants":["Stable Diffusion turns text and sketches into flexible floor plans","Text-to-layout model beats Pix2Pix and HouseDiffusion with LoRA","Cross-modal diffusion flexibly designs rooms from language and edges","Flexible residential layouts from text prompts via Stable Diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000661,"raw_usage":{"total_tokens":3028,"prompt_tokens":956,"completion_tokens":2072,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2000}},"tokens_in":572,"tokens_out":2072,"duration_ms":15994,"temperature":1.0,"reasoning_tokens":2000,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:06:34.863845+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the evaluation with the 8,000 comparison references drawn exclusively from a hold-out subset never used in LoRA fine-tuning, and report the four metrics separately for that unseen subset; if the LoRA-tuned model's advantage over HouseDiffusion and Pix2Pix shrinks substantially or reverses, the flexibility claim is falsified.","supporting_citations":[{"cited_title":"The LoRA network dimension was set to 32, undergoing training across 10 epochs with a batch size of 14","cited_arxiv_id":null,"evidence_quote":"The diffusion-based state-of-the-art baseline whose FID, PSNR, SSIM, and LPIPS scores the proposed method claims to beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The image-to-image GAN baseline that the paper compares against, representing the GAN-based approach it claims to surpass."},{"cited_title":"Mountstephens, J","cited_arxiv_id":null,"evidence_quote":"Supplies the RPLAN dataset and the graph-to-floor-plan method that the proposed natural-language representation extends and contrasts with."},{"cited_title":"As illustrated in Fig","cited_arxiv_id":null,"evidence_quote":"A graph-constrained layout generation method that represents the earlier single-input graph-based approach the paper critiques."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The graph-constrained GAN method with relational constraints, used as a prior-art anchor for design-rule representation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A prior diffusion-based multi-conditional floor-plan generator that the paper positions against to highlight its flexibility in input types."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior work evaluating text-to-image tools for floor plans, which motivates using natural language as the design-constraint modality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A graph-constrained high-rise residential floor-plan generation method from the same domain that the proposed approach extends."}],"review_version":1}