{"id":"86cfe4ba-87dd-4a1f-a7b4-5d1b1f342fe9","arxiv_id":"2607.23491","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Incomplete sketches refined by coarse-to-fine diffusion into vector floor plans, then agent-furnished inside verified boundaries, beat full-spec 2D baselines and Holodeck on rationality.","lead":"PlanCraft turns incomplete residential sketches into furnished 3D scenes by first completing a valid 2D floor plan, then placing furniture inside those fixed room boundaries. It matters because current 3D layout systems often skip that contract and produce overlapping or unusable rooms.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The \"25% sketch beats all fully specified baselines\" headline rests on sketches that are exact random subsets of the ground-truth plan being scored against — an information asymmetry, not just a realism gap, that may manufacture the margin.","rationale":"My concern shares the reader's root (SDP partial sketches are synthetic proxies) but sharpens it in a direction the reader did not fully articulate: even setting realism aside, the 25%-vs-baselines comparison is structurally uneven because the condition is drawn from the evaluation target itself. I mark \"partial\" agreement rather than full because the reader framed the risk purely as external-validity transfer, whereas I argue the headline comparison is internally stacked too. I do not adjust the verdict: CONDITIONAL already captures the right posture, since (a) the core 2D result (Tab. 1, graph-plus-block row, FID 4.026) does not depend on this concern, (b) the ablations (Tab. 2) are within-system and unaffected, and (c) the proposed human-sketch evaluation is exactly the kind of release/validation artifact the reader's conditions already gesture at. The expert-study and Holodeck-mapping caveats the reader noted stand independently. If the human-sketch test were run and failed, the progressive-design claim — the paper's first contribution — would need re-scoping, which would then justify a weaker verdict; but as submitted, CONDITIONAL with an explicit human-sketch validation condition is the honest position.","tokens_in":12725,"tokens_out":1819,"duration_ms":41003,"concrete_test":"Have a cohort of architects (e.g., the same 26 study participants) draw genuine ~25%-completeness sketches — rough strokes, no access to ground truth — for a sample of the RPLAN evaluation prompts, rasterize them, run PlanCraft-Diff, and recompute FID/IoU against Tab. 1 baselines. If FID rises above HouseDiffusion's 10.35 (or IoU drops below 0.412), the \"25% sketch outperforms all fully specified baselines\" claim does not transfer to real progressive inputs and must be re-scoped to synthetic vector subsets. A cheaper intermediate step: report PlanCraft-Diff metrics when the 25% condition is jittered (Gaussian-perturbed coordinates of the true elements) to measure sensitivity to sketch exactness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader flagged that SDP's synthetic progressions may not resemble real architect sketches. I think the issue is sharper and more load-bearing: the partial vectors are not merely stylized sketches, they are exact elements sampled from the same ground-truth plan the output is later scored against (Experiments, \"our automated pipeline progressively masks structural elements from complete plans\"). At 25% completeness the model receives a random quarter of the target's true walls/doors/windows at exact ground-truth coordinates, while every baseline in Tab. 1 receives only a graph or text prompt — a far less informative condition. FID/IoU/PSNR/SSIM are then computed against that same ground truth. So \"even at 25% completeness, PlanCraft surpasses all fully specified baselines\" compares a partial-answer-conditioned model against answer-blind models; the claim conflates input informativeness with progressive-design capability. This is compounded by the realism problem: an architect's early sketch is rough, imprecise geometry carrying intent, not a noiseless subset of the final drawing, so even the within-system Fig. 3 curve may not transfer. Two consequences: (1) the cross-method claim in Fig. 3/Tab. 1 is not apples-to-apples regardless of sketch realism; (2) the 61.1% FID headline is unaffected (that row uses the graph-plus-block condition), so the concern is localized to the completeness claims — but those are the paper's most distinctive selling point. Nothing here is internally inconsistent; it is a comparison-design risk.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper presents PlanCraft, a four-stage system that generates furnished 3D residential scenes from incomplete design sketches or natural-language programs. SketchPlan constructs partial-sketch training pairs from 80K RPLAN floor plans by masking structural elements under prerequisite constraints; PlanCraft-Diff is a modality-gated conditional DDPM with two-stage coarse-to-fine training and connectivity-based candidate filtering that refines incomplete inputs into complete floor plans; a rule-based post-processor vectorizes the raster output; and PlanCraft-Agent furnishes the verified vector plan using an LLM constraint generator plus a collision-aware solver. The authors report a 61.1% FID improvement over the best 2D baseline (HouseDiffusion), monotonic quality gains with sketch completeness (Fig. 3), ablations on backbone/training/filtering (Tab. 2), and a +15-point expert-rated Rationality advantage over Holodeck in a 26-participant architect study. The headline framing claims a 25%-complete sketch already outperforms all fully specified baselines.","tokens_in":13161,"tokens_out":2772,"duration_ms":39268,"significance":"If the results hold, this is a useful contribution: it is, to my knowledge, the first unified pipeline connecting progressive sketch completion to furnished 3D scene generation, and the \"floor plan as spatial contract\" argument is well-motivated and convincingly illustrated by the Holodeck/LLM failure cases in Fig. 6. Concrete strengths: a new 80K-plan partial-sketch dataset (SketchPlan) with a principled prerequisite-gated construction; a clean modality-gated DDPM design that shares parameters across two input types; ablations isolating backbone, curriculum, and filtering (Tab. 2); and an expert evaluation with 26 architecture-trained raters rather than crowdworkers. The graph-plus-block FID result (4.026 vs. 10.35 for HouseDiffusion) is large and, on its face, a genuine system-level result. The work is reproducible in outline: standard RPLAN data, standard DDPM/UNet, and full hyperparameters given.","major_comments":[{"comment":"The claim 'even at 25% completeness, PlanCraft surpasses all fully specified baselines' (Abstract; Experiments, 'Effect of Vector Completeness'; Fig. 3) compares conditions of unequal informativeness. Per the Experiments setup ('our automated pipeline progressively masks structural elements from complete plans'), the 25% vector input is an exact, noiseless subset of the ground-truth plan's walls/doors/windows at ground-truth coordinates — i.e., a quarter of the answer, unperturbed. Every Tab. 1 baseline instead receives only a graph or text prompt. FID/IoU/PSNR/SSIM are then computed against that same ground truth, so the comparison conflates input informativeness with progressive-design capability. This is the paper's most distinctive selling point and it is not apples-to-apples. Two fixes, either of which would suffice: (a) reframe the claim as an input-informativeness result and drop","section":"Fig. 3 / Abstract / Effect of Vector Completeness"},{"comment":"SDP simulates progressive drafting as exact subsets of the final vector plan with randomized order. A real architect's early sketch is rough, imprecise geometry encoding intent (approximate proportions, displaced walls), not a noiseless prefix of the final drawing. The paper motivates the entire system as 'architect-inspired' (Abstract; Introduction; Method, SDP paragraph), but no experiment tests robustness to geometric noise in the sketch (jittered vertices, approximate room extents), and no real or hand-drawn sketches are evaluated. As written, the Fig. 3 completeness curve characterizes performance on a synthetic proxy that may not transfer to the inputs the system claims to serve. At minimum, add a robustness study perturbing the partial vectors (vertex noise, room-block displacement at increasing sigma) and report metric degradation; a small qualitative study on actual architect sk","section":"Method (SDP) / Experiments setup"},{"comment":"The connectivity filter discards candidates whose floor-plan region is not a single connected component, and this alone moves FID from 11.674 to 4.026 — the single largest effect in the paper and most of the claimed 61.1% margin over HouseDiffusion (10.35). The manuscript does not state the evaluation protocol precisely: are Tab. 1 metrics computed only on the filtered subset of Ns=12 candidates? If so, FID is measured on a quality-curated sample, while baseline numbers (taken from prior publications or reproduced without filtering) are not given the same selection benefit, which would inflate the system-level comparison. The claim that filtering is 'a reliability filter instead of a hidden source of geometric improvement' is asserted but the IoU/PSNR/SSIM rows in Tab. 2 do move (0.498→0.510, etc.), so it is not purely distributional. Please (i) state exactly which samples enter each rep","section":"Tab. 2 / Method, Inference and Quality Filtering"},{"comment":"The user study (n=26) reports mean scores on three dimensions with a +15 Rationality gap, but the main text gives no variability or inferential statistics (no standard deviations, CIs, or significance tests), no statement on whether raters were blinded to system identity, and no discussion of ordering effects in the within-subjects design (each participant used both systems on 20 prompt-matched tasks). Given that PlanCraft's room boundaries are visually verifiable while Holodeck's failure modes are conspicuous (Fig. 6a), non-blind rating could inflate the gap. Please report per-dimension variance and a paired test, describe blinding and task randomization (even if in the supplement), and state whether participants had any affiliation with the project.","section":"Expert Evaluation on 3D Scene Generation"}],"minor_comments":[{"comment":"Including Holodeck in Tab. 1 (a 2D floor-plan metric table) is questionable framing: Holodeck is a 3D system whose internal layout was never designed to be scored as a floor plan, and mapping its vectors to RPLAN labels stacks the deck. Consider moving this row to a footnote or clearly labeling it as diagnostic.","section":"Tab. 1"},{"comment":"Tab. 2's annotation '↓65.5%' for the full model is computed against the one-stage DDPM, while the text elsewhere quotes the two-stage gain as 15.5% (13.810→11.674) — consistent, but the mixed reference points will confuse readers; please make each percentage's baseline explicit.","section":"Tab. 2"},{"comment":"Several free parameters are not ablated anywhere visible: the SDP wall-sampling probability λ, perturbation scales σ_blur and σ_block, the epoch split E1/E2=1000/500, and candidate count Ns=12. A short sensitivity table in the supplement would address reproducibility concerns.","section":"Method / Implementation Details"},{"comment":"All four Tab. 1 metrics are computed on 64×64 rasterized images. Since HouseDiffusion and HouseGAN++ are vector methods, rasterization resolution may differentially penalize them; please state how baseline outputs were rasterized and justify H=64, or report a vector-space metric (e.g., the authors' own metrics from Zeng et al. 2025, which is cited but not used).","section":"Evaluation Metrics"},{"comment":"The N-graph node position p_i ~ U(bbox_i) is underspecified: is a fresh random position sampled per training pair, and does this stochasticity in the condition affect evaluation determinism? A sentence of clarification suffices.","section":"Method, N-Graph construction"},{"comment":"Related Work omits discussion of sketch-to-image/layout completion literature (e.g., scribble- or partial-layout-conditioned generation) outside residential floor plans; positioning against partial-conditioning methods in adjacent domains would clarify novelty of the modality-gated design.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"The system's engineering is solid and the dataset contribution is real, but the two most quotable claims in the abstract — '25% sketch beats all fully specified baselines' and the architect-inspired framing — are exactly the two claims the current experiments do not support: the first rests on an information-asymmetric comparison, and the second on synthetic sketches that are noiseless subsets of ground truth. Both are fixable (reframing plus a noise-robustness study and possibly a matched-information baseline), so I recommend major revision rather than rejection. Also worth noting: the Floorplan-LLaMA baseline is the authors' own prior system, which is fine but should be kept in mind when weighing the baseline selection."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core here is not another text-to-layout demo. They couple partial-sketch completion to a verified 2D vector contract, then furnish inside those boundaries. That framing matches how residential design actually proceeds, and it is a cleaner split than Holodeck-style LLM layout hallucination.\n\nWhat is actually new: SketchPlan’s drafting-replay pairs on RPLAN (~80K), a single gated DDPM that takes either graph-plus-block or partial vectors with a completeness signal, coarse-to-fine training, connectivity filtering, and an agent that places furniture under explicit room polygons. Table 1’s graph-plus-block row is the clean result—FID 4.0 vs HouseDiffusion 10.4 (~61% lower), with IoU/PSNR/SSIM all up. Table 2’s ablations move in the right direction (two-stage, filtering). The expert study (n=26, architecture-trained) giving +15 Rationality over Holodeck is credible for a systems paper, even without error bars. Citations cover the right 2D and 3D baselines.\n\nSoft spot, and it is load-bearing for their distinctive claim: the “25% sketch already beats all fully specified baselines” comparison is not apples-to-apples. Those partial inputs are exact structural elements masked from the same ground-truth plan later used for FID/IoU scoring, at true coordinates. Baselines get only graph or text. That is partial-answer conditioning versus answer-blind conditioning, not a pure test of progressive design skill. Separately, SDP is randomized prerequisite-gated subsets, not rough architect strokes, so Fig. 3’s transfer to real early sketches is unproven. The graph-plus-block FID headline does not inherit this flaw; the completeness marketing does. Minor: no code/data release stated; expert stats lack significance tests; Holodeck’s mapped 2D metrics look harsh but the 3D expert gap still stands on its own.\n\nMath is standard conditional DDPM—nothing exotic, nothing broken. Data construction is reproducible in principle from RPLAN.\n\nWho it is for: people building residential layout tools or multi-room scene generators. Worth a serious referee. I would engage, cite the spatial-contract + progressive-conditioning idea, and push them to rerun completeness claims against noisy/approximate sketches and matched-information baselines.","headline":"Solid progressive floor-plan + grounded furnishing system; the 61% FID win is real, but the “25% sketch beats all baselines” line is an unfair comparison and should not be the headline.","tokens_in":14270,"tokens_out":603,"would_cite":true,"duration_ms":19958,"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":"Incomplete design sketches plus a fixed 2D floor-plan contract beat fully specified baselines at both 2D layout and furnished 3D residential scenes.","keywords":["floor plan generation","progressive design","diffusion models","3D indoor scene generation","spatial priors","residential layout","sketch conditioning"],"falsifier":"Give practicing architects real mid-design hand sketches (not algorithmically masked finished plans), run PlanCraft and the strongest fully specified baselines on the same briefs, and check whether the 25%-sketch FID and expert rationality advantages still hold.","tokens_in":13980,"feed_emoji":"🏠","tokens_out":836,"duration_ms":21404,"temperature":0.7,"pith_summary":"Automated home layout tools usually demand a finished graph or text brief up front and often skip a real floor plan when building 3D interiors. The authors argue that real design starts from rough strokes and that a locked 2D plan is the only reliable spatial contract before furniture is placed. They train on partial sketches derived from 80K real plans, refine those sketches into precise vector plans with a coarse-to-fine diffusion model, then place furniture inside the verified rooms. The result is much closer to real plan distributions than prior 2D generators and scores higher with architects on spatial rationality than language-model-only 3D systems—even when the input sketch is only one-quarter complete.","feed_headline":"25% sketches beat full briefs at home layout and 3D scenes","feed_subtitle":"A locked 2D floor plan as spatial contract cuts bad rooms and lifts architect-rated rationality by 15 points.","key_machinery":"SketchPlan plus PlanCraft-Diff: an automated replay of drafting that yields τ-labeled partial sketches, fed to a conditional diffusion model with modality-specific entry paths, two-stage coarse-to-fine training, and connectivity filtering; the verified plan then anchors PlanCraft-Agent’s constraint-based furniture placement.","core_discovery":"Design is progressive and the finished 2D floor plan is an irreplaceable spatial prior. A system that learns from partial sketches at every completeness level, sharpens them into geometrically valid vector plans, and only then furnishes inside those boundaries produces more faithful 2D layouts and more rational 3D scenes than methods that require fully specified inputs or that let language models invent layout and furniture together.","pith_inferences":["The same progressive-sketch plus spatial-contract pattern may transfer to multi-floor buildings, offices, or site plans where language-only layout also fails.","If real architect sketches differ systematically from the wall-first masking order, a small human-sketch fine-tune set could be the highest-leverage next experiment.","Editability via language-updated vectors or CAD round-trip suggests a practical human-in-the-loop loop rather than one-shot generation."],"forward_implications":["Floor-plan generators can accept rough or partial strokes instead of complete adjacency graphs or detailed text.","A verified 2D plan should sit between language and 3D assembly; skipping it produces overlapping rooms and bad proportions.","Sparse sketches (~25% complete) already suffice to beat fully specified prior methods on standard layout metrics.","Coarse-to-fine diffusion plus connectivity filtering is a practical recipe for closed, aligned residential polygons.","Expert-rated spatial rationality in furnished multi-room scenes improves when furniture is solved inside locked room boundaries."],"fun_headline_variants":["25% complete sketches beat fully specified home-layout baselines","Partial sketches sharpened to vector plans lift 3D rationality 15 points","Locked 2D floor plan as spatial contract cuts invalid rooms","PlanCraft replays architect strokes then furnishes inside fixed boundaries","Coarse-to-fine sketch refinement yields 61.1% lower floor-plan FID"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The automated step-by-step masking of walls then doors and windows on finished plans is assumed to stand in for the incomplete sketches architects actually draw, so scores at 25–100% completeness transfer to real progressive design.","fun_headline_variants_meta":{"raw":{"variants":["25% complete sketches beat fully specified home-layout baselines","Partial sketches sharpened to vector plans lift 3D rationality 15 points","Locked 2D floor plan as spatial contract cuts invalid rooms","PlanCraft replays architect strokes then furnishes inside fixed boundaries","Coarse-to-fine sketch refinement yields 61.1% lower floor-plan FID"]},"model":"grok-4.5","effort":"low","cost_usd":0.003405,"raw_usage":{"total_tokens":1198,"prompt_tokens":837,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":34048000,"prompt_tokens_details":{"text_tokens":837,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":284,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":837,"tokens_out":77,"duration_ms":6248,"temperature":1.0,"reasoning_tokens":284,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T20:55:50.546374+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Give practicing architects real mid-design hand sketches (not algorithmically masked finished plans), run PlanCraft and the strongest fully specified baselines on the same briefs, and check whether the 25%-sketch FID and expert rationality advantages still hold.","supporting_citations":[],"review_version":1}