{"id":"3d886a50-2880-46dc-b38c-aed1f96e50c6","arxiv_id":"2607.08086","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Modeling rooms as isotropic Gaussians and using them to initialize and guide diffusion yields controllable, editable polygonal floor plans that beat prior methods on RPLAN similarity and constraint metrics.","lead":"GRE-Diff generates and edits apartment floor plans by representing each room as a simple Gaussian blob (center plus size) that steers a diffusion model into clean polygons. It matters because it unifies constraint-aware generation with interactive LLM/GUI editing in one model, aiming at practical design tools.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Isotropic GRE may be insufficient as the room-level geometric state that steers diffusion, especially for non-rectilinear and edited layouts.","rationale":"The reader correctly isolates the isotropic-GRE sufficiency assumption as the weakest link and already flags the same limitation evidence (Sec. 4.5 / Fig. 8) and the useful but incomplete Table S1 ablation. That concern is load-bearing for the strongest claim: the reported SOTA numbers and the unified generation+editing story are presented as consequences of GRE steering diffusion. Because the paper never compares against a richer room-level geometric state, we cannot yet tell whether the gains come from the continuous probabilistic idea itself or from the particular isotropic parameterization that happens to fit RPLAN’s mostly Manhattan apartments. The concrete covariance/box ablation would settle the issue without requiring external data or code release. No stronger internal inconsistency appears; the rest of the pipeline (dual-path encoder, autoregressive GRE prediction, anchor-mask editing) is coherent given that representation. Therefore the verdict remains CONDITIONAL with the same high-confidence caveats the reader listed (no code, asymmetric baselines, non-expert user study, acknowledged brittleness).","tokens_in":23692,"tokens_out":696,"duration_ms":7456,"concrete_test":"On the same RPLAN split, retrain/evaluate an otherwise identical GRE-Diff variant that replaces isotropic σ with a full 2\times2 covariance (or oriented axis-aligned box) inside GuidanceNet and the μ t/σ t schedules of the reverse process. Report FID/KID/MMD/COV/BC/RC/F1 and the fraction of Fig. 8-style failures (non-rectilinear boundaries, connectivity, uneven vertices). If the richer state does not improve metrics by >5% relative or cut irregular-boundary failures by a clear margin, the isotropic-GRE sufficiency claim holds; otherwise the load-bearing representation assumption weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on GRE (isotropic N(μi, σi^{2}I) with only centroid and scalar scale; Sec. 3.1) being a sufficient continuous room-level state that both initializes and parameterizes the reverse process (Eqs. 1–3 / reverse update in Sec. 3.2 and Supp. C.2). The paper itself shows failures precisely where this approximation is weakest: irregular/non-Manhattan boundaries, uneven vertex density, connectivity errors, and semantic mismatches (Sec. 4.5, Fig. 8). The ablation that most supports the claim (Table S1) only contrasts full GRE-state vs. conditioning-only vs. plain DDPM; it does not test whether a richer room state (anisotropic covariance, oriented box, or multi-mode blob) would improve the same metrics or reduce the documented failure modes. If isotropic GRE is under-expressive for real residential polygons, the headline gains (FID 4.36, COV 96.09%, F1 BC-RC 99.21%) and the “unified gen+edit” story are partly artifacts of RPLAN’s rectilinear bias rather than a generally sufficient representation.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes GRE-Diff, a diffusion framework for controllable apartment floor-plan generation and interactive editing. Each room is represented by an isotropic Gaussian Room Embedding (GRE) with centroid μ and scalar scale σ; GuidanceNet predicts these embeddings from multi-modal conditions (room types/counts, boundary, optional polygons), and DenoisingNet denoises polygons initialized and steered by the GRE-parameterized reverse process. The system supports LLM-parsed natural-language constraints and five GUI editing operators (Add/Delete/Move/Repurpose/Anchor) with hard fixed-room masks. On RPLAN, GRE-Diff reports best or near-best FID/KID/MMD, highest COV, and highest F1 BC-RC among seven baselines, plus strong editing Tiny-ROE/F1, user preference over Graph2Plan and GT, and high overlap-free/reachability rates. Ablations contrast dual-path attention and GRE-as-state vs conditioning-only vs plain DDPM.","tokens_in":24036,"tokens_out":1269,"duration_ms":20700,"significance":"If the results hold under fair comparison, this is a useful systems contribution: a single diffusion model that unifies vectorized generation and localized editing under boundary and room-count constraints, with a practical dual-mode (LLM/GUI) interface. Strengths include multi-metric evaluation on a held-out RPLAN split, an explicit ablation showing that wiring GRE into the diffusion state transition (not merely as external conditioning) improves metrics (Table S1), functionality rates (overlap-free 95.56%, reachability 96.35%), a user preference study, and an honest limitations section with failure cases. The continuous room-level state idea is a reasonable alternative to ordered vertex sequences and discrete masks for structured layout diffusion.","major_comments":[{"comment":"Sec. 3.1–3.2 and Supp. C.2: The central representation claim is that an isotropic GRE N(μi, σi²I) is a sufficient room-level geometric state to initialize and parameterize polygon diffusion (forward/reverse Eqs. 1–3). Table S1 only contrasts full GRE-state vs conditioning-only vs plain DDPM; it does not test richer room states (anisotropic covariance, oriented box, multi-mode blob). Sec. 4.5 and Fig. 8 document failures precisely where isotropy is weakest (non-rectilinear/irregular boundaries, uneven vertex density, connectivity/semantics). Either add a controlled ablation with a richer room state under the same training budget, or substantially qualify the claim that GRE is a generally adequate geometric state rather than an effective prior for RPLAN’s largely rectilinear apartments.","section":null},{"comment":"Table 1 and Sec. 4.2: Baseline fairness needs tighter documentation. HouseDiffusion is evaluated without boundary constraints by design; Graph2Plan’s strong numbers depend on retrieval-assisted post-processing (FID rises to 20.10 and RC falls to 50.19% without it). WallPLAN/GSDiff use snapping/alignment. Please report a primary comparison under a common protocol (same boundary+room-type conditions, no non-neural post-processing, or with post-processing clearly factored out for all methods) so that the headline ranking (FID 4.36, F1 BC-RC 99.21%) is not partly an artifact of unequal pipelines.","section":null},{"comment":"Sec. 4.1–4.2 functionality metrics and Supp. on doors/windows: Overlap-free and reachability rates (95.56% / 96.35%) use annotated door labels to build connectivity graphs, while the method itself does not generate doors/windows and instead applies a post-hoc rule-based procedure (Supp. limitation 4). Clarify whether reachability is measured on model outputs alone or after rule-based opening insertion, and report connectivity without relying on GT door annotations if the claim is that generated layouts are functionally usable end-to-end.","section":null}],"minor_comments":[{"comment":"Fig. 7 caption refers to “Fig. S5” in the main text narrative (“Fig. S5 shows…”); renumber or fix the cross-reference so main-text figures are self-contained.","section":null},{"comment":"Notation: the reverse update in Sec. 3.2 uses αt / ᾱt inconsistently with Supp. C.2; unify the noise schedule symbols and define ᾱt explicitly in the main text.","section":null},{"comment":"Table 2 Tiny-ROE = 0.00% for all operations is striking; briefly state the LPIPS threshold and how “requested edit applied” is verified (semantic label check vs geometric IoU) so the metric is reproducible.","section":null},{"comment":"User study (Sec. 4.2): report inter-rater agreement or confidence intervals on the 1740 responses; preference over GT is interesting and would benefit from a short discussion of possible confounds (rendering style, wall thickness).","section":null},{"comment":"Related work: briefly position GRE against BlobGAN/BlobCtrl-style spatial Gaussians and PolyDiffuse set diffusion, which are cited but not contrasted on the room-state design choice.","section":null},{"comment":"Typos/clarity: “load-bearing wall constraints” in Fig. 5 caption is not fully explained in the main text; define how load-bearing walls enter the conditioning y.","section":null}],"recommendation":"minor_revision","confidential_remarks":"Solid empirical systems paper on a standard benchmark; suitable for a CV/graphics venue after tightening baseline protocol and GRE-sufficiency discussion. The isotropic-GRE concern is real but domain-appropriate for RPLAN; I would not reject on that alone if claims are scoped. No red flags on circularity or fabricated metrics from the provided evidence."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a solid systems paper in layout generation. The actual new piece is Gaussian Room Embedding—an isotropic 2D Gaussian (centroid + scalar scale) that both initializes and parameterizes the reverse diffusion of room polygons, not just conditions it. That, plus a dual-path GuidanceNet with gen/edit adapters and a unified set of edit operators (add/delete/move/anchor/repurpose), is what separates it from HouseDiffusion, Graph2Plan, MaskPLAN, and ChatHouseDiffusion.\n\nWhat it does well: the RPLAN numbers are clean and multi-axis (FID 4.36, COV 96%, F1 BC-RC 99%, editing Tiny-ROE 0% with high F1, overlap-free ~96%, reachability ~96%). The supp ablation that pits full GRE-state vs conditioning-only vs plain DDPM is the right experiment and lands in their favor. Limitations figure and text are honest about non-Manhattan failures. Citations cover the right prior art; the diffusion math is standard DDPM with room-dependent drift/variance, not hand-wavy.\n\nSoft spots, in proportion: isotropic GRE is thin for elongated, rotated, or multi-mode rooms, and Fig. 8 shows exactly where that bites (irregular contours, connectivity, semantics). That is a real representation limit, not a fatal one—the paper is mostly evaluated where RPLAN is rectilinear. No code/checkpoints. Some baseline comparisons are slightly asymmetric. User study is CS students, not architects. Free parameters exist (loss weights, bins K, γ) but nothing unusual for the genre.\n\nThe stress-test note is fair as a caveat and overstated as a sink: the paper never claims GRE is a universal geometric state; it claims it works for this task on this data, and the state-vs-conditioning ablation supports that claim. Who this is for: people building interactive CAD/layout tools or controllable structured diffusion. It is not a paradigm shift; it is a useful incremental method paper.\n\nI would send it to peer review. Engage if you work on floor plans or layout diffusion; otherwise skim GRE + Table S1 and move on.","headline":"Competent interactive floor-plan diffusion paper: GRE as a room-level state is a real, ablated idea, metrics on RPLAN are strong, and the isotropic limitation is real but already partly owned by the authors.","tokens_in":24645,"tokens_out":561,"would_cite":false,"duration_ms":17078,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Modeling each room as a simple Gaussian lets one diffusion model both generate and edit apartment floor plans under user constraints.","keywords":["Floor Plan","Layout Generation","Controllable Editing","Diffusion Models","Gaussian Room Embedding","Interactive Design","Polygonal Layouts"],"falsifier":"Run the model on a held-out collection of strongly non-rectilinear apartment outlines never seen in RPLAN; if FID, overlap-free rate or door-reachability collapse while room-count accuracy remains high, the isotropic GRE state is insufficient.","tokens_in":24571,"feed_emoji":"🏠","tokens_out":874,"duration_ms":19531,"temperature":0.7,"pith_summary":"Apartment floor-plan design forces a designer to explore a huge space of room arrangements while keeping boundaries valid, rooms non-overlapping, and functional relationships intact. This paper claims that representing every room by an isotropic Gaussian—its center plus one scale number for extent—supplies a continuous room-level state that both initializes and steers a diffusion process. The resulting GRE-Diff system turns natural-language room counts or GUI sketches into diverse polygonal layouts, then supports local add/delete/move/anchor edits without regenerating everything from scratch. On the RPLAN residential dataset the method reports the best overall distributional match, highest coverage, and highest constraint satisfaction among the compared approaches. The practical goal is a single controllable pipeline that keeps AI suggestions and human refinement inside the same generative loop.","feed_headline":"Gaussians for rooms let AI generate and edit floor plans","feed_subtitle":"One continuous embedding per room unifies constraint-aware creation and local refinement on real apartments.","key_machinery":"Gaussian Room Embedding (GRE): each room is a spatial Gaussian N(μ, σ²I); the predicted (μ, σ) set the initial noise and the room-dependent drift/variance of the diffusion, after which DenoisingNet recovers explicit polygons.","core_discovery":"The paper establishes that Gaussian Room Embeddings—each room cast as an isotropic two-dimensional Gaussian whose mean is the centroid and whose scalar variance captures spatial extent—form a continuous, order-robust geometric state that can initialize and parameterize a conditional diffusion process, thereby unifying high-quality polygonal floor-plan generation with structure-preserving interactive editing under boundary, room-type, and user-edit constraints.","pith_inferences":["If GRE already unifies generation and editing, the same room-level state could serve as a shared interface for multi-turn agentic design loops that plan, check and re-edit layouts.","Failures on highly irregular contours suggest that anisotropic or multi-mode Gaussians may be required before the method generalizes beyond residential Manhattan-like plans.","Because GRE is compact and order-invariant, it is a natural candidate for coupling with downstream furniture-placement modules to close the empty-room-to-furnished-interior pipeline.","The same continuous room state may transfer to multi-story or commercial layouts once training data cover those typologies."],"forward_implications":["Users can obtain diverse, constraint-satisfying apartment polygons from natural-language room counts and sketched boundaries without hand-built graphs.","The same diffusion model supports add, delete, move, repurpose and anchor edits while freezing user-chosen rooms and preserving global coherence.","Reported metrics on RPLAN show lower FID/KID/MMD, higher coverage and higher boundary-plus-room F1 than the compared graph, GAN and diffusion baselines.","Automatic BC/RC feedback can resample failing layouts without external snapping or post-processing.","A dual-mode interface (LLM instructions or GUI operators) lets non-experts iterate designs inside one generative loop."],"fun_headline_variants":["Rooms as Gaussians drive constrained floor-plan diffusion","Gaussian embeddings unify floor-plan generation and edits","Continuous room Gaussians enable interactive layout diffusion","GRE casts rooms as Gaussians for editable polygonal plans","Isotropic Gaussians power constraint-aware floor-plan diffusion"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"A single center point and one size number per room are enough to steer diffusion into valid, functional apartment polygons even when boundaries are irregular or rooms are edited.","fun_headline_variants_meta":{"raw":{"variants":["Rooms as Gaussians drive constrained floor-plan diffusion","Gaussian embeddings unify floor-plan generation and edits","Continuous room Gaussians enable interactive layout diffusion","GRE casts rooms as Gaussians for editable polygonal plans","Isotropic Gaussians power constraint-aware floor-plan diffusion"]},"model":"grok-4.5","effort":"low","cost_usd":0.004546,"raw_usage":{"total_tokens":1311,"prompt_tokens":734,"num_sources_used":0,"completion_tokens":81,"cost_in_usd_ticks":45460000,"prompt_tokens_details":{"text_tokens":734,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":496,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":734,"tokens_out":81,"duration_ms":4572,"temperature":1.0,"reasoning_tokens":496,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T13:19:39.586444+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the model on a held-out collection of strongly non-rectilinear apartment outlines never seen in RPLAN; if FID, overlap-free rate or door-reachability collapse while room-count accuracy remains high, the isotropic GRE state is insufficient.","supporting_citations":[],"review_version":1}