{"id":"905295b1-8b60-44d7-a8e8-ba3e99ead852","arxiv_id":"2607.03990","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"InSpace generates complete structure-aware 3D indoor scenes (layout plus textured assets) from a single equirectangular 360° image via three-stage flow matching with view- and asset-selective attention.","lead":"InSpace turns one 360° panoramic photo into a full 3D indoor room with walls, floors, and separately textured furniture. It matters because layout-aware scenes are needed for robotics, digital twins, and simulation, where ordinary single-view methods leave objects floating or intersecting walls.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The central claim rests on a synthetic benchmark and a tiny, manually-aligned baseline comparison; real-world and fair-comparison evidence is too thin to fully support the outperformance claim.","rationale":"The paper is a solid engineering contribution: view-selective and asset-selective attention are well-motivated, the ablations (Table 1, Fig. 10) cleanly isolate their benefit, and ERP-FRONT is a useful resource. The reader correctly flags the monocular-depth prior as a soft spot, yet that prior is only one component; the more load-bearing weakness for the *central claim as stated* is the evaluation design that underwrites “outperforming \times generalizing.” Because the comparative numbers rest on 20 manually-aligned synthetic crops and real-world transfer is essentially uncontested, the claim remains conditional on stronger, automatic, full-scene, and real-data evidence. This does not overturn the reader’s CONDITIONAL verdict; it simply relocates the primary risk from Stage-1 geometry to the comparative evaluation protocol. No change of verdict category is required, only a clearer emphasis on what still needs to be shown.","tokens_in":26332,"tokens_out":585,"duration_ms":5798,"concrete_test":"Re-run the full Stage-2/3 pipeline on the 20 comparison scenes using automatic ICP-only alignment (no manual pre-alignment) and report full-scene (not crop-only) CD/F1/IoU for InSpace vs. SAM3D; if InSpace’s margin collapses below the Table D.1 gap or layout IoU falls below the ERP-FRONT numbers, the outperformance claim weakens. Separately, add any available real panoramic baseline on the 7 ReplicaPano rooms.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim asserts that InSpace produces complete layout+assets scenes that outperform prior single-image generators on 3D/2D metrics and generalizes to ReplicaPano. The quantitative backbone of that claim is almost entirely ERP-FRONT (synthetic, same distribution as training). The only head-to-head numbers against SceneGen/MIDI/SAM3D use 20 scenes, require manual alignment + ICP (Sec. 6.2, Supp. D), and evaluate only the visible perspective crop rather than full-scene layout. ReplicaPano evidence is qualitative plus a 7-scene self-reconstruction table (Supp. C.3) with no competing methods. If the synthetic advantage or the manual-alignment protocol is the main driver of the reported gains, the outperformance and generalization parts of the claim do not hold outside the authors’ own distribution. The reader’s depth/PSG concern is real but secondary: even perfect depth would not rescue a claim whose comparative evidence is this narrow.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"InSpace generates complete 3D indoor scenes—structural layout (floors/walls) plus separately textured assets—from a single equirectangular (ERP) 360° image. The pipeline has three stages: (1) monocular depth lifting to a Partial Scene Geometry (PSG) and calibrated camera center; (2) coarse dense-voxel structure generation with view-selective cross-attention over cubemap features, plus optional Layout-Guided Structure Inversion from the PSG and a 3D OBB detector; (3) layout- and asset-aware fine geometry/texture via global self-attention and asset-selective cross-attention under flow matching, built on TRELLIS.2 O-Voxel latents. The authors introduce ERP-FRONT (~29K synthetic ERP–mesh pairs from 3D-FRONT) and report 3D (IoU, CD, F1) and 2D (PSNR, LPIPS) metrics, ablations of view-selective attention and inversion, and limited comparisons to SceneGen, MIDI, and SAM3D plus qualitative/self-reconstruction results on ReplicaPano.","tokens_in":26666,"tokens_out":1236,"duration_ms":14273,"significance":"The problem framing is well motivated: limited-FOV single-image scene generators omit structural layout and often produce floating or interpenetrating assets. Using full 360° ERP context as the spatial anchor, together with view-selective and asset-selective attention that couple voxel latents to geometrically visible image regions, is a clear technical step beyond asset-only multi-instance pipelines. ERP-FRONT is a useful community resource for paired ERP-to-scene training. Ablations cleanly isolate view-selective attention (Stage-2 IoU ~44→57, CD roughly halved) and inversion (t0 sweeps), which strengthens confidence in the core mechanisms. If the comparative and real-world evidence were tightened, this would be a solid contribution to structure-aware indoor scene generation.","major_comments":[{"comment":"Sec. 6.2 and Supp. D: Head-to-head comparison with SceneGen, MIDI, and SAM3D uses only 20 scenes, requires manual alignment then ICP, and evaluates asset-level metrics only inside a perspective crop (red dashed boxes), not full-scene layout. This protocol is too narrow and operator-dependent to support a general outperformance claim for complete layout+asset scenes. Either expand to a larger automatic-alignment protocol (or report full-scene metrics where baselines can be fairly scored), or clearly restrict claims to the ERP-FRONT self-benchmark and treat baselines as qualitative context.","section":null},{"comment":"Supp. C.2–C.3 (ReplicaPano): Generalization is shown on 7 scenes with self-reconstruction metrics and no competing methods. Domain-gap failures (frames, curtains, texture style) are acknowledged, but the abstract/conclusion still state that InSpace “generalizes well” to realistic panoramas. Strengthen this with more scenes, failure-rate reporting, and/or a limited baseline comparison under the same ERP input; otherwise tone the generalization claim to match the evidence.","section":null},{"comment":"Sec. 4.1–4.2, Eq. (5), Fig. 3: View-selective masks and Layout-Guided Structure Inversion both depend on monocular depth and the calibrated camera center c. There is no sensitivity study to depth error, camera-height error, or PSG incompleteness (e.g., wrong depth → wrong visibility cones and inversion prior). A controlled noise/ablation on depth or c would show whether the Stage-2 gains remain when the spatial prior is imperfect—especially important for real ERP inputs where depth is noisier than on ERP-FRONT.","section":null}],"minor_comments":[{"comment":"Table 1: Optimal t0 differs between Stage 2 (0.5) and Stage 3 (0.7); the text explains this, but a single recommended default and a short note on how users should choose t0 would help reproducibility.","section":null},{"comment":"Fig. 2 and Sec. 1: “Existing methods” are illustrated with floating/misplacement; cite which specific outputs (method + figure) are shown so the failure modes are attributable.","section":null},{"comment":"Sec. 4.2: Cubemap FoV=120° and α=50 are fixed; a brief sensitivity note (or appendix) would clarify robustness of M_vs.","section":null},{"comment":"Supp. B / Table B.1: 3D OBB F1@0.75 is low (~29%); discuss how OBB size/yaw error propagates into Stage-3 asset quality, since Stage 3 is conditioned on these boxes.","section":null},{"comment":"Related Work: PanoContext-Former is noted as closest prior with unavailable code; a short qualitative discussion of its reported limitations versus InSpace’s outputs would better position novelty.","section":null},{"comment":"Notation: N0/s0 vs 16^3 latent resolution appears in several places; keep a single consistent symbol for the coarse latent grid size.","section":null}],"recommendation":"major_revision","confidential_remarks":"The technical core (view-selective conditioning + layout/asset decomposition on sparse voxels) is credible and the ERP-FRONT resource is valuable. The skeptic’s concern about thin comparative and real-world evidence is well founded and is the main reason I recommend major rather than minor revision; I do not see internal circularity or a broken derivation. Fit for a solid CV venue is good if comparisons and generalization claims are brought in line with the evidence. No novelty-disclosure or citation-pattern issues stood out."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is that InSpace is a competent three-stage pipeline that finally treats floors/walls as first-class geometry rather than an afterthought, and it comes with a new paired ERP-to-mesh dataset (ERP-FRONT) that the community can actually use. That combination is worth knowing about.\n\nWhat is new is concrete: view-selective cross-attention that routes each voxel only to the cubemap faces visible from a calibrated camera center, asset-selective ROI masking for the fine stage, Layout-Guided Structure Inversion that seeds the coarse latent from the monocular depth prior, and the 26.5k/2.5k ERP-FRONT split built on 3D-FRONT. The ablations cleanly isolate the view-selective mask (IoU 44\to57, CD roughly halved) and the inversion t0 sweep. The method is built on TRELLIS.2 sparse voxels and flow matching, so the math and training objective are standard and look solid; citations are ordinary background rather than circular. Qualitative results on both synthetic rooms and ReplicaPano look coherent, and the layout-plus-assets decomposition is a real practical advantage over MIDI/SceneGen/SAM3D-style asset-only pipelines.\n\nSoft spots are real but proportionate. The head-to-head numbers against the three single-image baselines use only 20 scenes, require manual ICP alignment, and evaluate only the visible crop rather than full-scene layout. ReplicaPano is seven scenes of self-reconstruction with no competitors. Depth-derived PSG is a load-bearing prior; if it is systematically off, both the visibility mask and inversion degrade. No code or weights are released. These do not sink the synthetic results, but they keep the outperformance and generalization claims provisional.\n\nThis is for people building indoor digital twins, robotics simulators, or panoramic 3D content pipelines. It deserves a serious referee. I would engage with the work, cite the dataset and the attention design if I am in the area, and expect the paper to improve under review once the comparison protocol is tightened.","headline":"Solid engineering paper that correctly targets the missing layout problem in indoor scene generation and ships a useful dataset; comparative claims rest on thin real-world and baseline evidence.","tokens_in":27312,"tokens_out":512,"would_cite":true,"duration_ms":5369,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A single 360° image is enough to generate a complete indoor 3D scene with floors, walls, and separately textured furniture.","keywords":["3D scene generation","equirectangular projection","indoor layout","flow matching","view-selective attention","sparse voxels","ERP-FRONT"],"falsifier":"Replace the estimated depth with systematically biased or noisy depth on the same ERP inputs; if view-selective attention and layout-guided inversion still produce coherent rooms whose IoU and asset placement match the reported numbers, the claim that the geometric prior is essential fails.","tokens_in":27209,"feed_emoji":"🏠","tokens_out":595,"duration_ms":5244,"temperature":0.7,"pith_summary":"Single-image 3D generators already make good individual objects, but indoor scenes need the room itself—floors, walls, and the geometry that anchors every piece of furniture. A narrow camera view does not supply enough spatial coverage for that layout, so assets float, collide, or sit in the wrong place. InSpace takes one equirectangular 360° image instead, estimates a partial point cloud and camera center, then runs a three-stage flow-matching pipeline that first builds a coarse voxel room structure with view-selective attention and then fills in detailed layout and per-asset geometry and texture with a hybrid global-local attention. The authors also release ERP-FRONT, a large paired ERP-to-mesh dataset built from 3D-FRONT. The result is a complete, textured indoor scene whose layout and assets are spatially coherent, outperforming prior single-image scene generators on both geometric and rendering metrics and generalizing to real panoramic scans.","feed_headline":"One 360° photo yields a full 3D room with layout and furniture","feed_subtitle":"View-selective attention turns a panorama into floors, walls, and textured assets that stay grounded","key_machinery":"View-selective cross-attention (and its asset-selective counterpart): each voxel or asset token is allowed to attend only to the cubemap faces or image regions that are geometrically visible from the calibrated camera center, so the model composes the full room without receiving conflicting cues from opposite walls.","core_discovery":"From a single equirectangular 360° image, InSpace recovers a complete 3D indoor scene that includes both the structural layout (floors and walls) and separately textured assets, with placements that respect the recovered geometry, by cascading partial-geometry priors, view-selective coarse structure generation, and hybrid global-local refinement under flow matching.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Single 360° ERP image yields full 3D room layout plus textured assets","InSpace rebuilds floors walls and grounded furniture from one panorama","Structure-aware cascade turns equirectangular photo into complete indoor 3D","Partial geometry then hybrid attention: 3D scene from solitary 360° view","View-selective flow matching recovers layout and assets from one ERP image"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The monocular depth map and the camera center derived from it must be accurate enough that the resulting visibility masks and partial geometry really do point each voxel at the correct faces of the panorama.","fun_headline_variants_meta":{"raw":{"variants":["Single 360° ERP image yields full 3D room layout plus textured assets","InSpace rebuilds floors walls and grounded furniture from one panorama","Structure-aware cascade turns equirectangular photo into complete indoor 3D","Partial geometry then hybrid attention: 3D scene from solitary 360° view","View-selective flow matching recovers layout and assets from one ERP image"]},"model":"grok-4.5","effort":"low","cost_usd":0.005512,"raw_usage":{"total_tokens":1512,"prompt_tokens":794,"num_sources_used":0,"completion_tokens":100,"cost_in_usd_ticks":55120000,"prompt_tokens_details":{"text_tokens":794,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":618,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":794,"tokens_out":100,"duration_ms":5152,"temperature":1.0,"reasoning_tokens":618,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T22:28:24.501398+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Replace the estimated depth with systematically biased or noisy depth on the same ERP inputs; if view-selective attention and layout-guided inversion still produce coherent rooms whose IoU and asset placement match the reported numbers, the claim that the geometric prior is essential fails.","supporting_citations":[],"review_version":1}