{"id":"ede71ac5-385c-4504-b8b2-1680a58b5235","arxiv_id":"2411.12279","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A two-stage system where an LLM creates an initial floorplan layout from text and a conditional diffusion model refines it is claimed to beat existing floorplan generators on RPlan.","lead":"HouseTune generates floorplans from natural-language descriptions in two stages: a large language model sketches a rough layout, then a diffusion model refines it into a final floorplan. It reports improved scores over prior methods on the RPlan dataset, but the paper's evaluation has unresolved inconsistencies and no released code.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) defines an invalid diffusion process: the forward marginal at T is shifted by e(y) while the reverse chain starts from N(0,I), so the training loss and sampling update are mutually inconsistent.","rationale":"Among the reader's concerns, the strongest is the inconsistency between the conditioned forward process and the unconditioned reverse chain. The empirical issues (missing code, no error bars, LLM-generated evaluation text) cast doubt on the reported numbers, but a mathematical flaw in the core diffusion formulation is more fundamental: even if the experiments are reproduced exactly as described, the method as written does not implement a valid conditional diffusion model. The added e(y) term in Eq. (4) shifts the forward marginal at every timestep; at T the latent is centered at e(y) rather than at zero. The reverse chain in Eq. (6) ignores this shift, so the generative distribution will not match the data distribution conditional on y. This breaks the theoretical foundation of the 'dual-conditioning' contribution and makes the state-of-the-art claim unexplained. The paper deserves credit for a plausible two-stage pipeline and for comparing against strong baselines, but the central methodological equation needs correction or a proper derivation. A minimal synthetic experiment would settle whether the inconsistency is real or whether the implementation silently uses a different reparameterization.","tokens_in":11964,"tokens_out":3205,"duration_ms":34580,"concrete_test":"Implement the exact forward process of Eq. (4) on a simple 2D Gaussian mixture with a fixed condition y, train with the loss in Eq. (10), and generate samples via Eq. (6) starting from N(0,I). Compare the empirical marginal distribution of the generated samples to the true conditional data distribution; if the mean differs by e(y) or an accumulated shift, the proposed reverse process is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (4) defines xt = sqrt(alpha_t)*x0 + sqrt(1-alpha_t)*epsilon + e(y). Thus q(xt|x0) = N(sqrt(alpha_t)*x0 + e(y), (1-alpha_t)I). At t=T, sqrt(alpha_T) is approximately 0, so q(x_T) is approximately N(e(y), I). The reverse process in Eq. (6) starts from x_T ~ N(0,I) and applies the standard DDPM update, which is derived for a forward process with mean sqrt(alpha_t)*x0 and variance (1-alpha_t)I. With the added e(y) term, the posterior q(x_{t-1}|x_t,x_0) is no longer the one used by Eq. (6), and the noise-prediction loss in Eq. (10) is not a valid variational lower bound. Unless e(y) is absorbed by reparameterizing x0 (e.g., x_t = sqrt(alpha_t)*(x0+e(y)) + sqrt(1-alpha_t)*epsilon) or the reverse chain is initialized at x_T ~ N(e(y), I) with a corrected update, the model does not sample from the intended conditional distribution. This undermines the central claim that the dual-conditioning diffusion refinement achieves state-of-the-art results, because the proposed mechanism is mathematically inconsistent as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HouseTune, a two-stage text-to-floorplan generation framework. In the first stage, a multimodal LLM (GPT-4o) with a Chain-of-Thought prompt and a small set of demos produces an initial layout (Layout-Init) in JSON from a natural-language description. In the second stage, a conditional diffusion model refines Layout-Init into a final vector floorplan (Layout-Final), with conditioning applied both in the forward noising process and the reverse denoising process. The authors evaluate on the RPlan dataset against House-GAN, House-GAN++, HouseDiffusion, PuzzleFusion, and Tell2Design, reporting improvements in realism, diversity, and compatibility, plus ablations on one- versus two-stage generation, prompt design, LLM choice, and conditioning placement. The abstract claims state-of-the-art performance across all metrics.","tokens_in":12236,"tokens_out":5954,"duration_ms":57013,"significance":"If the method were sound, it would provide a practical text-to-floorplan pipeline that reduces reliance on domain-specific labeled training data and leverages LLM common-sense reasoning. The two-stage decomposition and the CoT prompting strategy are appealing and could inspire follow-up work. However, the central diffusion formulation is mathematically inconsistent as presented, and the experimental reporting has several gaps (missing appendix, no error bars, contradictory table prose). These issues currently prevent the paper from supporting its central claims.","major_comments":[{"comment":"Equation (4) defines the forward process as xt = sqrt(alpha_t)*x0 + sqrt(1-alpha_t)*epsilon + e(y), so the marginal is q(xt|x0) = N(sqrt(alpha_t)*x0 + e(y), (1-alpha_t)I). At t=T this is approximately N(e(y), I). The reverse process described in Section 3.2.1 starts from a fully noised sample xT and applies Equation (6), which is the standard DDPM update derived for a forward process without the additive e(y) term. No posterior q(x_{t-1}|x_t,x0) is derived for the modified process, and the reverse chain is not initialized at N(e(y), I). Consequently, the noise-prediction loss in Equation (10) is not a valid variational bound for the sampling procedure, and the model is trained and sampled under inconsistent distributions. This is a load-bearing flaw: the claim of dual-conditioning diffusion refinement and the state-of-the-art results rest on a mechanism that is not mathematically well-defined as stated. The authors must reparameterize e(y) into the signal (e.g., xt = sqrt(alpha_t)(x0 + e(y)) + sqrt(1-alpha_t)*epsilon) or derive the corrected posterior and reverse initialization, then retrain and re-evaluate.","section":"3.2.1"},{"comment":"Table 3b and the surrounding text contradict each other. The prose states \"as the conditional ratio increases, model performance gradually declines,\" but the table shows Macro IoU and Micro IoU increasing monotonically with the rate: 18.67%/15.43% at 1e-3, 20.12%/16.59% at 1e-2, and 21.37%/18.46% at 1e-1. Since the arrows in the table denote higher-is-better, the table implies the opposite conclusion. This matters because the conditional participation ratio is one of the method's free parameters, and the recommended setting is unclear from the present text.","section":"4.3"},{"comment":"The paper repeatedly refers to an Appendix for network architecture, implementation details, and the Text-to-Layout baseline (e.g., Section 3.2.2, Section 4.1, Section 4.2), but the arXiv v4 manuscript contains no appendix. The statement \"Our code will be made publicly available\" is not accompanied by a link or repository identifier. These omissions prevent reproducibility and make it impossible to verify the training setup or the one-stage baseline.","section":"4.1"},{"comment":"Tables 1 through 4 report no error bars, standard deviations, or significance tests. The headline claims, such as the 28% diversity improvement over HouseDiffusion (8.6 vs. 11.2 on task 5) and the 79% compatibility improvement, are based on single runs. Given the stochastic nature of both the LLM (temperature sampling) and the diffusion process, the absence of variance estimates does not support the abstract's claim of state-of-the-art performance across all metrics.","section":"Tables 1-4"}],"minor_comments":[{"comment":"The numbers for HouseTune in the \"unlabeled data\" comparison (Micro IoU 21.84%, Macro IoU 17.75%) and in the \"Training on Artificial Instructions Only\" comparison (Micro IoU 15.69%, Macro IoU 11.43%) appear inconsistent; the paper should clarify which setting each number corresponds to.","section":"4.2"},{"comment":"Equation (11) uses symbols C^s and C^t that are not defined in the text; presumably these denote the predicted and ground-truth corner coordinates, but this should be stated explicitly.","section":"3.2.3"},{"comment":"The caption contains a typo: \"Layou-Final\" should be \"Layout-Final.\"","section":"Figure 1"},{"comment":"The reference \"What Makes In-Context Learning Work. Rethinking the role of demonstrations: ...\" appears to be an incomplete citation; the authors should provide the full bibliographic entry.","section":"References"},{"comment":"The claim of a 79.88% diversity improvement over Tell2Design is not directly readable from Table 1 as formatted, because the table's column structure for Diversity across task sizes is ambiguous; the authors should present the comparison more clearly.","section":"4.2"}],"recommendation":"reject","confidential_remarks":"The central diffusion formulation in Eq. (4)-(6) is mathematically inconsistent, and this is not a local presentation issue: it affects the validity of the training loss, the sampling procedure, and therefore the reported state-of-the-art numbers. The missing appendix and lack of statistical rigor further weaken the manuscript. These issues would require a substantial rewrite and re-validation, which is beyond a minor or major revision in the usual sense. I recommend rejection, though the two-stage idea and the CoT prompting design may be worth pursuing in future work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Frankly, this one is a mixed bag. The two-stage idea—LLM with CoT writes a rough layout, then a diffusion model cleans it up with conditioning on that layout in both forward and reverse—is genuinely new and worth discussing. I haven't seen that exact combination in the floorplan literature. The qualitative results look sensible, and the prompt ablations with different LLMs are a useful sanity check.\n\nBut the central math doesn't hold up as written. Equation (4) defines the forward process as x_t = sqrt(alpha_t) x0 + sqrt(1-alpha_t) eps + e(y), so the marginal at time T is approximately N(e(y), I), not N(0,I). Yet the reverse process starts from x_T ~ N(0,I) and uses the standard DDPM update (6), which is only derived for the unshifted forward. The paper never computes the posterior q(x_{t-1}|x_t,x0) for the shifted process, so the noise-prediction loss in (10) and the sampling update are mutually inconsistent. This is not a cosmetic issue; it's the core mechanism. Either e(y) needs to be absorbed into a reparameterized x0 or the reverse chain and posterior need to be re-derived. As it stands, the paper does not actually describe a valid diffusion model.\n\nThe experimental section has its own problems. Table 3b shows that decreasing the conditional rate from 1e-1 to 1e-3 decreases both Macro and Micro IoU, but the text says 'as the conditional ratio increases, model performance gradually declines.' That's backwards. There are no error bars or significance tests anywhere, the code and appendix are promised but absent, and the evaluation uses GPT-4o to generate both the test descriptions and the Layout-Init conditions, so the loop is partially closed.\n\nNone of this means the idea is dead. With a corrected derivation, a re-run of the ablations, and some statistical rigor, this could become a solid paper. But as submitted, I can't take the SOTA claim seriously.\n\nMy recommendation: this deserves a serious referee, but with the expectation of heavy revision. If you're looking for a cautionary example of why conditioning in diffusion needs to be done carefully, this is a good one.","headline":"A novel two-stage LLM-plus-diffusion idea undermined by an invalid forward-process derivation and contradictory ablation results.","tokens_in":12774,"tokens_out":4178,"would_cite":false,"duration_ms":43063,"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 two-stage pipeline — an LLM drafts a rough layout, then a dual-conditioned diffusion model refines it — beats prior floorplan generators on realism, diversity, and compatibility.","keywords":["text-to-floorplan generation","large language models","chain-of-thought prompting","conditional diffusion models","Layout-Init","dual conditioning","RPlan dataset"],"falsifier":"Re-run the pipeline with standard DDPM forward noising while keeping reverse-only conditioning and compare FID and compatibility; if the scores match the reported values, the dual-conditioning mechanism in Eq. (4) is not the source of the gains. The alternative is to derive the posterior $q(x_{t-1}\\mid x_t, x_0)$ for the modified forward process and check whether it equals the update used in Eq. (6).","tokens_in":11779,"feed_emoji":"🏠","tokens_out":12871,"duration_ms":113370,"temperature":0.7,"pith_summary":"The paper is trying to establish that a floorplan generator does not need a large corpus of labeled layout graphs or exact geometric annotations. Instead, an LLM uses chain-of-thought prompting to turn a natural-language request into a rough initial layout, and a conditional diffusion model refines that sketch into a final floorplan. The claim is that this division of labor makes generation more user-friendly and matches or beats specialized GAN and diffusion baselines on the RPlan benchmark, with especially large gains in diversity and compatibility.","feed_headline":"Beats prior floorplan generators with LLM sketch and diffusion polish","feed_subtitle":"LLM common sense supplies the layout sketch; a dual-conditioned diffusion model refines it into a final floorplan.","key_machinery":"The load-bearing mechanism is the dual-conditioned noise process: the forward noising step adds the learned Layout-Init embedding $e(y)$ to every noisy sample, and the reverse update keeps using that embedding while predicting the denoised layout. The reverse network is a Transformer with three structured-masking attention types (component-wise self-attention, global self-attention, relational cross-attention), a continuous decoder for coordinate noise, and a discrete 'int2bit' decoder that rounds coordinates to 8-bit binary values to preserve collinearity and orthogonality.","core_discovery":"The central discovery is that conditioning a diffusion model on an LLM-produced Layout-Init in both the forward noising process and the reverse denoising process gives better final floorplans than conditioning only during denoising as HouseDiffusion does. The forward process injects the Layout-Init embedding $e(y)$ into every noisy sample via $x_t = \\sqrt{\\alpha_t}x_0 + \\sqrt{1-\\alpha_t}\\epsilon + e(y)$, and the reverse process uses the same embedding when predicting the denoised layout. On RPlan, this two-stage recipe yields the best reported realism, FID-based diversity, and graph-edit compatibility scores across room counts 5 to 8, including a 28% diversity improvement and a 79% compatibility improvement over HouseDiffusion.","pith_inferences":["The success of this recipe suggests that any domain with a cheap, coarse spatial prior — furniture arrangement, warehouse layout, schematic diagrams — could adopt the same LLM-sketch-then-diffusion-refine pattern.","An apples-to-apples benchmark in which HouseTune and Tell2Design receive identical plain-text prompts and identical room-count constraints would separate the benefit of the two-stage architecture from the difference in supervision.","Because the paper reports that conditioning ratio strongly affects FID, a practical follow-up is to anneal the conditioning strength during training or sampling rather than keep it fixed."],"forward_implications":["If the claim holds, text-to-floorplan systems can be built without large labeled layout-graph datasets: the LLM supplies the structural prior and the diffusion model handles geometric refinement.","The dual-conditioning recipe is transferable: conditioning the forward noising process as well as the reverse process should improve other conditional diffusion tasks where a rough layout or structural prior is available.","Natural-language interfaces become viable for non-expert users, since room counts, types, and approximate adjacency can be specified in plain text rather than bubble diagrams or exact geometry.","The method's consistency across GPT-4o, DeepSeek-R1, and Doubao-1.5Pro suggests the prompting strategy, not the specific LLM, carries the first stage.","The ablation results indicate that injecting too much conditional information degrades generation quality, so conditioning strength is a tunable knob rather than a fixed design choice."],"supporting_citations":[{"why":"Supplies the HouseDiffusion base architecture and discrete/continuous decoders that HouseTune adapts, and is the primary baseline for diversity and compatibility comparisons.","marker":"Shabani et al. [2023]"},{"why":"Defines the Chain-of-Thought prompting strategy that produces Layout-Init from the LLM.","marker":"Wei et al. [2022]"},{"why":"Provides the DDPM formulation and the masked L2 noise-prediction loss used in training.","marker":"Ho et al. [2020]"},{"why":"Supplies the RPlan dataset and the room-count splits used for all experiments.","marker":"Wu et al. [2019]"},{"why":"Provides the Tell2Design text-to-layout dataset and baseline used for the language-guided comparison.","marker":"Leng et al. [2023]"},{"why":"Defines the House-GAN baseline and the Realism/Diversity/Compatibility evaluation procedure that HouseTune inherits.","marker":"Nauata et al. [2020]"},{"why":"Defines FID, the metric used to measure diversity in Table 1.","marker":"Heusel et al. [2017]"},{"why":"Defines exact graph edit distance, used for the Compatibility score.","marker":"Abu-Aisheh et al. [2015]"}],"fun_headline_variants":["LLM sketch + dual-conditioned diffusion: floorplan SOTA by wide margin","HouseTune two-stage: LLM common sense, diffusion precision","Text-to-floorplan: LLM plans, diffusion nails details","Dual conditioning on LLM layout: +28% diversity, +79% compatibility"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that adding the Layout-Init embedding directly into every noisy sample still behaves like a standard diffusion process, so that starting from ordinary random noise and reversing the noising steps stays valid; if that consistency is broken, the reported refinement gains collapse.","fun_headline_variants_meta":{"raw":{"variants":["LLM sketch + dual-conditioned diffusion: floorplan SOTA by wide margin","HouseTune two-stage: LLM common sense, diffusion precision","Text-to-floorplan: LLM plans, diffusion nails details","Dual conditioning on LLM layout: +28% diversity, +79% compatibility"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001487,"raw_usage":{"total_tokens":5930,"prompt_tokens":865,"completion_tokens":5065,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":4984}},"tokens_in":481,"tokens_out":5065,"duration_ms":38732,"temperature":1.0,"reasoning_tokens":4984,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:43:47.582227+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the pipeline with standard DDPM forward noising while keeping reverse-only conditioning and compare FID and compatibility; if the scores match the reported values, the dual-conditioning mechanism in Eq. (4) is not the source of the gains. The alternative is to derive the posterior $q(x_{t-1}\\mid x_t, x_0)$ for the modified forward process and check whether it equals the update used in Eq. (6).","supporting_citations":[{"cited_title":"House-gan: Relational generative adversarial networks for graph-constrained house layout generation","cited_arxiv_id":null,"evidence_quote":"Defines the House-GAN baseline and the Realism/Diversity/Compatibility evaluation procedure that HouseTune inherits."},{"cited_title":"An exact graph edit distance algorithm for solving pattern recognition problems","cited_arxiv_id":null,"evidence_quote":"Defines exact graph edit distance, used for the Compatibility score."}],"review_version":1}