{"id":"6506add4-28a5-414a-9a9b-d567d502db75","arxiv_id":"2412.06859","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Fine-tuning a latent diffusion model on 500 footprint-text-floorplan triples yields plausible floorplans for several building types, though the quantitative evaluation is weak and irreproducible.","lead":"An AI model trained on 500 paired examples takes a building outline and a short text description and draws a plausible floorplan for stadiums, offices, apartments, libraries, and auditoriums. It matters because it tests whether general image AI can help architects and non-experts explore design options quickly, but the paper's own quality metrics are oddly weak.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No quantitative evidence that generated floorplans respect the input footprint or text; if the ControlNet condition is ignored, the central conditional-generation claim collapses.","rationale":"The reader's weakest assumption identified exactly the condition-adherence problem: the paper provides no quantitative check that generated layouts respect the footprint mask or the textual program. My stress-test agrees and sharpens the concern by noting that the reported metrics are marginal, not conditional: FID, KID, SSIM, and PSNR compare generated images to real images globally, so a model that ignores both conditions could still score well if it produces plausible floorplans. The small dataset (500 pairs) increases the risk of the ControlNet adapter being underfit or collapsed. The human evaluation also does not resolve this, since it rates overall design quality rather than whether the output matches the specified footprint and brief; the statistically significant gap (p=0.001) is described as 'close,' which is an internal inconsistency. Because the central claim is conditional generation, the absence of any conditional fidelity metric is a load-bearing gap. The proposed concrete test—occupancy IoU against the input mask plus a mask-swap control—would directly settle whether the footprint condition matters, and a prompt-swap classification test would settle whether the text condition matters. I do not see a reason to move the verdict beyond the reader's CONDITIONAL: the qualitative figures suggest the method is plausible, but the evidence as presented is insufficient to accept the strong claim of 'superior performance across all metrics' or of reliable multi-condition control.","tokens_in":12546,"tokens_out":3204,"duration_ms":29474,"concrete_test":"On a held-out split of the 500 pairs, generate N=20 samples per test input. Measure spatial consistency: binarize each generated floorplan to an occupancy mask, resize to the footprint mask resolution, and report IoU between the generated occupancy and the input mask, plus the fraction of generated non-background pixels lying outside the mask. Compare against two baselines: (a) the same text-conditioned LDM without ControlNet, and (b) a control where the footprint mask is replaced by a different footprint for the same prompt. If the model's IoU is not significantly higher than baseline (a), or if outputs do not change when the mask changes, the footprint condition is not functioning. For text adherence, fix one footprint, vary only the prompt across building types, and have architects classify the generated building type; chance-level accuracy would show the text condition is ignored.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is conditional generation p(x|y1,y2), where y2 is the footprint mask and y1 is the text brief (Sec. 3.1). The load-bearing assumption is that the ControlNet stage actually uses y2 and y1 to shape the output. The evaluation in Sec. 4 never tests this joint dependency: Table 1 reports FID, KID, SSIM, and PSNR between generated and real images, which are marginal distribution metrics and can be excellent even if conditions are ignored (e.g., a model that always emits plausible generic floorplans). The human game (Table 2) shows real floorplans rated 6.89 vs. generated 5.36 with p=0.001, which the text calls 'close' despite being statistically significant; this does not establish conditional adherence. The paper itself notes in Sec. 4 (Evaluation Metrics) that producing multiple samples for identical inputs makes quantitative evaluation challenging, yet it provides no conditional fidelity metric. With only 500 training pairs, a plausible failure mode is that the frozen LDM dominates and the ControlNet adapter learns a trivial mapping, so outputs correlate weakly with mask and prompt. Figure 6 shows visual variety, but variety alone is consistent with ignoring both conditions and sampling from the prior. Thus the central claim—that the model translates a given footprint and brief into a matching floorplan—is unsupported by the reported evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a latent diffusion model for generating architectural floorplans conditioned on two inputs: a building footprint mask (image) and a textual design brief. The method follows the ControlNet recipe: a frozen text-to-image LDM is augmented with a trainable copy controlled by the footprint mask, and the two-stage objective is the standard denoising loss with conditions y1 (text) and y2 (footprint). The authors introduce a new dataset of 500 (mask, prompt, floorplan) triplets, report FID/KID/SSIM/PSNR comparisons against LDM and DALL-E 2, present a blinded human rating game with architects, visualize denoising behavior and latent space structure, and claim a scale-agnostic capability across building types such as stadiums, offices, apartments, libraries, and auditoriums.","tokens_in":12805,"tokens_out":3296,"duration_ms":33300,"significance":"If the conditional-generation claim is solid, the work would be a useful step toward practical AI-assisted architectural design: it addresses a real gap (most floorplan generators are residential-only and require expert zoning inputs), contributes a novel multimodal dataset, and proposes a gamified human-evaluation protocol that is more appropriate for design quality than generic image metrics. The blinded human game is a genuine strength, as is the use of an external frozen foundation model. However, the current evidence does not establish that the model actually respects the footprint or the text prompt, and several quantitative claims are internally inconsistent. The significance will depend on whether the authors can supply conditional-fidelity measurements and correct the evaluation framing.","major_comments":[{"comment":"The reported quantitative metrics are both internally implausible and not diagnostic of conditioning. SSIM=0.130 (on a scale where 1.0 is identical) and PSNR=7.596 dB are far below values normally described as 'excellent image quality' (typically SSIM>0.9 and PSNR>30 for reconstruction tasks); the text in 'Model Performance' should be corrected. More importantly, FID, KID, SSIM, and PSNR compare marginal image distributions between generated and real floorplans; they can be excellent even if the model ignores y1 and y2 and simply samples plausible generic floorplans. The paper needs conditional fidelity metrics, for example overlap between the generated layout and the input footprint mask, consistency of room labels with the requested program, or a correlation measure between varied prompts and outputs while holding the footprint fixed.","section":"Section 4, Table 1"},{"comment":"The human evaluation results contradict the text's characterization. Real floorplans scored 6.89 vs. generated 5.36 with p=0.001 and t=3.917; this is a statistically significant gap, not 'slightly notable differences' or 'closeness of the scores' as stated in 'Model Performance' and the caption of Figure 5. Furthermore, the game asked architects to rate 'composition and architectural integrity' only, so it does not test whether the generated image follows the given footprint or design brief. A conditional adherence task (e.g., 'does the layout fit the provided outline?' or 'does it include the requested room types?') is needed to support the central p(x|y1,y2) claim.","section":"Section 4, Table 2 and Figure 5"},{"comment":"Training on only 500 author-created triplets with batch size 1 raises a real risk that the frozen LDM dominates and the ControlNet adapter learns a trivial or partially ignored mapping from y2. The paper shows training/validation losses in Figure 4, but no split details and no condition-dropout or input-randomization ablations are reported. A concrete test would be to shuffle the footprint masks (or prompts) against the training pairs and measure whether generated outputs change accordingly; alternatively, report metrics such as mask-conditioned FID or per-prompt layout similarity to demonstrate that the conditions are load-bearing.","section":"Section 4, 'Model Performance' and Section 6.1"},{"comment":"The scale-agnostic claim is only supported by qualitative figures. The text states the model 'seamlessly adapt to diverse scales' and 'relies on a ratio-based driven approach,' but no operational definition of scale-agnosticism or quantitative test is given. I suggest fixing the footprint while varying the intended building type (or vice versa), then measuring properties such as internal room proportions, corridor widths, or the footprint-occupancy ratio to show that the model adapts its output to the input conditions rather than producing a generic layout.","section":"Section 4, 'Scale-agnostic approach' and Figure 6"},{"comment":"The baseline comparison is not sufficient to support the paper's superiority claims. LDM and DALL-E 2 are text-to-image models that do not receive the footprint condition y2, so Table 1 conflates conditioning ability with generic image quality; a model that outputs plausible images without respecting the footprint could still win on FID. To isolate the contribution of the ControlNet stage, the authors should compare against conditional baselines (e.g., pix2pix-style footprint-to-floorplan translation, ControlNet without text conditioning, or graph-based floorplan generators) under the same evaluation protocol.","section":"Section 4, 'Comparing the results to baselines' and Table 1"}],"minor_comments":[{"comment":"The list of contributions numbers two items as '3)'; renumber the contributions.","section":"Section 1"},{"comment":"There is a typo 'residdential'; should be 'residential'.","section":"Section 4, 'Datasets'"},{"comment":"The text references 'Figure 15' for human evaluation results, but the figure appearing in the main text is Figure 5; the supplementary material also contains a Figure 15, creating confusion. Please renumber or disambiguate.","section":"Section 4, 'Model Performance'"},{"comment":"The FloorplanGame link is missing a URL; provide an accessible link or a detailed description of the game interface in the supplementary material.","section":"Section 4, 'Model Performance'"},{"comment":"There is a discrepancy in training epochs: the main text says '549 training cycles (epochs)' while the supplementary says '429 epochs'. Please reconcile.","section":"Section 6.1"},{"comment":"Minor typos: 'monitoried' should be 'monitored', 'transfrom' should be 'transform', and 'createdd' should be 'created'.","section":"Section 6.1"},{"comment":"The notation F(x; Θ) is ambiguous: x is described as the input image, but in Eq. (4) x also appears as the argument inside Z(y2; Θz1). Clarify the role of x (the original image versus the latent) in the residual connection.","section":"Section 3.1, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The central idea is potentially valuable, but the review found no evidence that the generated floorplans respect the footprint and prompt conditions; the paper's own quantitative results (especially Table 1 and Table 2) are inconsistently reported. These issues are fixable with additional experiments within the manuscript's scope, so major revision rather than rejection seems appropriate. I would also ask the editor to verify that the dataset's provenance and labeling process are described in sufficient detail for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a genuinely useful new asset—a 500-sample dataset of floorplan images paired with footprint masks and text briefs across several building types—and the qualitative figures suggest the LDM+ControlNet pipeline works after a fashion. But the evaluation section does not measure what the title promises: there is no quantitative check that the generated floorplan actually respects the input footprint or the text program. The stress-test note is right, and it lands on the paper's load-bearing claim.\n\nWhat is new: the dataset and the task definition. Prior floorplan generators are residential-only; this one conditions on a footprint mask plus a text brief and shows results for stadiums, offices, apartments, libraries, auditoriums. That is a real step for early-design exploration. The architecture is a direct application of LDM [44] with ControlNet [62]; no new generative machinery, but that is fine if the application is evaluated properly.\n\nSoft spots, in order of severity. First, no conditional-fidelity metric. Table 1 reports FID/KID/SSIM/PSNR between generated and real images, which only measure marginal distributions; a model that ignores both conditions and emits generic plausible floorplans could score well. The human game (Table 2) is blinded and compares against real floorplans, but it tests overall quality, not whether the output matches the given footprint or brief. The paper itself notes that multiple samples for identical inputs make quantitative evaluation challenging, which makes the absence of any overlap or program-consistency metric more glaring. Second, the numbers are internally mismatched: SSIM=0.130 and PSNR=7.596 are very low absolute values, yet the text calls them 'excellent'; the human scores differ with p=0.001, and the text calls that 'close.' Those need fixing or reframing, not just copyediting. Third, the promised ablation study (contribution 3) is missing from the experiments; there is a section on denoising steps and latent visualization, but no ablation of the conditioning mechanism. Fourth, the dataset and code are not released, so the central empirical contribution is currently unverifiable. Minor: training details conflict (549 epochs in Fig. 4 vs. 429 in supplementary), and the contribution list has a duplicated '3)'.\n\nNone of this is fatal to the underlying idea. The dataset and task are worth having, and a revised version with a proper conditional-fidelity evaluation (e.g., footprint overlap, room-label agreement) and released data would be a solid contribution. As it stands, the paper's claims outrun its evidence.\n\nFor peer review: send it to a serious referee. The task is important enough and the dataset is a real asset, but the referee should push hard on evaluation and require the ablation and data release.","headline":"A useful multi-type floorplan dataset and a plausible LDM+ControlNet baseline, but the evaluation never tests the conditional claim it leads with.","tokens_in":13313,"tokens_out":2040,"would_cite":false,"duration_ms":19843,"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":"A single latent diffusion model can generate floorplans for many building types from a footprint mask and a text brief, without being given the scale.","keywords":["floorplan generation","latent diffusion model","ControlNet","conditional image generation","architectural design","design brief","footprint conditioning","scale-agnostic"],"falsifier":"Compute the overlap between the generated floorplan and the input footprint mask (for instance, the intersection-over-union of the building outline), or check that changing the text prompt with a fixed footprint produces substantially different room layouts; if the mask overlap is no better than an unconditioned baseline, or the layout barely changes with the prompt, the central claim of condition-following generation would fail.","tokens_in":12351,"feed_emoji":"🏢","tokens_out":5173,"duration_ms":49147,"temperature":0.7,"pith_summary":"The paper aims to show that one generative model, built on latent diffusion, can take two simple inputs—a building's footprint mask and a text description of the desired program—and output a plausible architectural floorplan for a variety of building types, including apartments, offices, stadiums, libraries, and auditoriums. If this works, it would let architects and non-experts quickly explore design alternatives without manually constructing complex parametric rules or computing scale. The authors introduce a new dataset of 500 footprint–prompt–floorplan triples and a two-stage training scheme: a text-conditioned latent diffusion model is first trained on a large general image–text corpus, then a ControlNet-style adapter is trained on the new dataset to add footprint conditioning. They report that their model outperforms two baseline text-to-image systems on FID, KID, SSIM, and PSNR, and that architects rate its outputs close to real floorplans in a blind evaluation game.","feed_headline":"Diffusion model turns footprint and text into floorplans","feed_subtitle":"One model adapts to stadiums, offices, apartments, and libraries from a mask plus a one-line brief.","key_machinery":"The machinery is a two-stage latent diffusion architecture. The first stage is a text-conditioned latent diffusion model with cross-attention over tokenized text, pretrained on a large general image–text corpus. The second stage freezes that model, clones it, and connects the clone through zero-initialized convolution layers—a ControlNet-style adapter—so that the footprint mask is injected as a second conditioning signal. The training objective is the standard denoising loss $L = \\mathbb{E}_{z_0, t, y_1, y_2, \\epsilon}\\left[\\lVert \\epsilon - \\epsilon_\\theta(z_t, t, y_1, y_2)\\rVert_2^2\\right]$, where $y_1$ is the text prompt and $y_2$ is the footprint image; the zero convolutions let the adapter be added without disrupting the pretrained text-conditioned model.","core_discovery":"The central claim is that a conditional latent diffusion model can learn the mapping from a building footprint mask plus a text design brief to a floorplan, and that it does so in a scale-agnostic way—without the scale being provided explicitly. The authors support this with generated examples across several building functionalities and with quantitative metrics: their model achieves the lowest FID (22.436), the lowest KID (1.844), the highest SSIM (0.130), and the highest PSNR (7.596) when compared with two baseline text-to-image models. They also present a human evaluation in which architects gave generated floorplans an average score of 5.36 out of 10 versus 6.89 for real floorplans, a difference that is statistically significant (p = 0.001).","pith_inferences":["The claimed generality rests on a dataset of only 500 triples, so conditioning fidelity is unverified quantitatively; a dedicated test of whether the output respects the input mask and prompt would be needed.","Image-level metrics such as FID and SSIM do not measure whether the text and footprint are actually followed; a prompt-consistency or footprint-overlap metric would be more informative.","The human evaluation, while suggestive, uses a small sample of 30 images and 10 architects, so a larger blind study would be needed to confirm that generated floorplans are close to real ones in functional quality.","The latent-space visualization suggests the model learns functional proximity between building types, but this is qualitative; a quantitative clustering evaluation could strengthen the claim."],"forward_implications":["A single model can serve multiple building typologies, unlike earlier floorplan generators that focus almost exclusively on residential layouts.","Designers can explore alternative programs on the same footprint by changing only the text prompt, without recomputing scale or dimensions.","Non-experts could generate plausible floorplan starting points quickly, potentially accelerating early-stage architectural ideation.","The model's reported ability to fuse design elements across building types suggests it can produce configurations that are not simple copies of training examples."],"supporting_citations":[{"why":"Supplies the base latent diffusion model and the two-stage conditioning concept that the paper adapts for floorplan generation.","marker":"[44]"},{"why":"Provides the ControlNet-style zero-convolution adapter used to inject the footprint condition in the second stage.","marker":"[62]"},{"why":"Supplies the large text–image pretraining dataset from which the first-stage model is resumed.","marker":"[46]"},{"why":"Serves as a baseline text-to-image system that the paper compares against in Table 1.","marker":"[37]"},{"why":"Represents prior floorplan-generation work that motivates the need for a multi-functional dataset and conditioning setup.","marker":"[50]"}],"fun_headline_variants":["Latent diffusion model generates floorplans from text and footprint","One latent diffusion model adapts to stadiums, offices, and homes","Diffusion model turns footprint and text into floorplans for any building","Latent diffusion model generates novel floorplans for stadiums to homes","Architectural AI: latent diffusion turns footprint and brief into floorplans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model actually obeys the input footprint and text prompt rather than ignoring them and generating a generic floorplan; the paper shows only visual examples and gives no quantitative measure of conditioning adherence.","fun_headline_variants_meta":{"raw":{"variants":["Latent diffusion model generates floorplans from text and footprint","One latent diffusion model adapts to stadiums, offices, and homes","Diffusion model turns footprint and text into floorplans for any building","Latent diffusion model generates novel floorplans for stadiums to homes","Architectural AI: latent diffusion turns footprint and brief into floorplans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000765,"raw_usage":{"total_tokens":3382,"prompt_tokens":923,"completion_tokens":2459,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":2364}},"tokens_in":539,"tokens_out":2459,"duration_ms":16853,"temperature":1.0,"reasoning_tokens":2364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:58:40.735473+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the overlap between the generated floorplan and the input footprint mask (for instance, the intersection-over-union of the building outline), or check that changing the text prompt with a fixed footprint produces substantially different room layouts; if the mask overlap is no better than an unconditioned baseline, or the layout barely changes with the prompt, the central claim of condition-following generation would fail.","supporting_citations":[{"cited_title":"High-resolution image syn- thesis with latent diffusion models, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the base latent diffusion model and the two-stage conditioning concept that the paper adapts for floorplan generation."},{"cited_title":"Adding conditional control to text-to-image diffusion models, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the ControlNet-style zero-convolution adapter used to inject the footprint condition in the second stage."},{"cited_title":"Dall·e 2: Ai art generation","cited_arxiv_id":null,"evidence_quote":"Serves as a baseline text-to-image system that the paper compares against in Table 1."},{"cited_title":"FloorGAN: Generative network for automated floor layout generation","cited_arxiv_id":null,"evidence_quote":"Represents prior floorplan-generation work that motivates the need for a multi-functional dataset and conditioning setup."}],"review_version":1}