{"id":"8895f366-f51a-428f-a507-d1fe4d26504c","arxiv_id":"2607.16805","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-free pipeline combining occupancy-based view selection, early-stop sign-consensus velocity fusion, and rigid Gaussian layout refinement improves SAM3D-based 3D scene generation on Replica and ScanNet++.","lead":"This paper presents a training-free pipeline that extends the single-image 3D generator SAM3D to multi-view indoor scene generation by selecting complementary views, fusing early flow steps with sign consensus, and refining object poses with Gaussian rendering. It reports large scene-level Chamfer-distance gains on Replica and ScanNet++, though the headline numbers compare against its own pose-refinement variant rather than against prior methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Method silently requires per-pixel depth D_i (Eqs. 1, 7, 13) but states a calibrated-RGB problem; experiments use only dataset GT depth, so the 'calibrated multi-view' claim overreaches.","rationale":"The reader's weakest_assumption exactly identifies the depth/mask availability issue. I agree this is the most load-bearing concern: it challenges whether the method actually solves the problem it claims to solve (calibrated multi-view RGB scene generation) rather than merely being a depth-assisted pipeline validated on depth-rich datasets. The misreported percentage in the abstract (43.8%/30.9% as 'over SAM3D with ICP' when it is actually Ours-GS vs. Ours-ICP) is a real reporting error, but it is fixable and does not invalidate the core method; the depth gap is structural. The paper gives no depth estimator, uses no depth in the problem statement, yet requires depth in every major module. This is not an internal inconsistency in the equations, but an unstated input requirement that limits transfer. The proposed concrete test—using monocular depth in place of GT depth—would settle whether the method remains effective in the RGB-only regime. Since the reader already marked CONDITIONAL and the concern strengthens that conditionality without requiring outright rejection, the verdict should remain CONDITIONAL. I set verdict_should_be to UNCHANGED to indicate no change to the reader's verdict is needed; the paper must address this gap before acceptance.","tokens_in":17425,"tokens_out":9328,"duration_ms":92807,"concrete_test":"Run the full Scene-SAM3D pipeline on a Replica subset with depth maps replaced by monocular depth estimates from a pretrained network (e.g., Depth Anything V2) for all D_i in Eqs. 1/7/13, keeping all other settings fixed. Recompute instance-level CD/Comp and scene-level CD/IoU, and compare with the GT-depth results in Tables 1/2. If the scene-level reduction vs. MD+SAM3D+ICP shrinks substantially (e.g., below 20%) or the pipeline fails to produce coherent layouts, the paper's headline gains depend on ground-truth depth rather than an RGB-only solution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's problem definition (Sec. 2) lists observations as (I_i, M_i, C_i, K) with no depth map. Yet the proposed pipeline depends on depth throughout: Eq. 1 uses D_i to build world-space point sets for anchor/helper occupancy selection; Eq. 7 uses D_a to map anchor pixels into the canonical latent grid for sign-consensus regions; and Eq. 13 includes L_depth, a depth-supervision term in layout refinement. The authors never specify how D_i is obtained for arbitrary calibrated video. On Replica and ScanNet++ depth is read directly from dataset ground truth, so all reported instance- and scene-level results are conditional on having depth. Without D_i, the view-selection criterion (Eq. 2) is undefined and the Gaussian layout refinement loses one of its primary signals. The abstract's 'calibrated multi-view scene asset generation' implies RGB-only input, which the method does not actually support as described. This is a load-bearing gap: a practitioner applying Scene-SAM3D to typical RGB-only multi-view captures cannot execute the view-selection or layout-refinement steps without an external depth estimator that the paper neither specifies nor evaluates.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Scene-SAM3D proposes a training-free pipeline that adapts the frozen single-image generator SAM3D to calibrated multi-view scene asset generation. The method selects a compact anchor-helper view set via world-space occupancy coverage, fuses helper-view latent velocities only in early flow steps with an anchor-guided sign-consensus operator, and then refines the object-level rotations, scales, and translations through a rigid-object Gaussian layout optimization. Experiments on Replica and ScanNet++ report instance-level improvements over SAM3D and a MultiDiffusion-based baseline, and large scene-level CD reductions attributed to the Gaussian layout refinement. The paper also reports roughly 20% savings in flow-model FLOPs and latency under ten-view inference.","tokens_in":17711,"tokens_out":6330,"duration_ms":54196,"significance":"If the claims are reproducible, the paper offers a practically useful design pattern for lifting object-centric 3D generators to multi-view scenes: view selection by 3D occupancy, early-stop conflict-aware fusion, and layout refinement via a frozen Gaussian proxy are all sensible, and the component-wise ablations (Tables 1, 3, 5) support the internal design choices. However, the headline scene-level claim is not supported by the tables as currently worded, and the method's unstated dependence on per-pixel depth is a load-bearing gap. With corrections to the comparisons, explicit handling of depth, and proper statistical reporting, the contribution would be significant for the 3D generation community.","major_comments":[{"comment":"The abstract states that Scene-SAM3D reduces scene-level CD by 43.8% on Replica and 30.9% on ScanNet++ 'over SAM3D with ICP alignment.' In Tables 2 and 6, the rows compared are 'Ours ICP' and 'Ours GS' under the same coverage-selected views and fusion strategy. On Replica, (8.64-4.86)/8.64 = 0.4375; on ScanNet++, (6.89-4.76)/6.89 = 0.3091. Thus the percentages are an ablation of the GS layout refinement against the authors' own ICP baseline, not a comparison with SAM3D+ICP. The claim as worded is unsupported. Please either add the missing SAM3D+ICP comparison or change the claim to 'compared with our ICP-based refinement.'","section":"Abstract & Sec. 4.1, Tables 2, 6"},{"comment":"The problem definition lists observations as (I_i, M_i, C_i, K) with no depth map, and the paper frames the task as 'calibrated multi-view scene asset generation.' However, D_i is used in Eq. (1) to build world-space point sets for view selection, in Eq. (7) to map anchor pixels into the canonical latent grid, and in Eq. (13) as L_depth supervision in layout refinement. The source of D_i is never specified; on Replica and ScanNet++ it is the dataset ground-truth depth. Without a depth estimator, a practitioner with RGB-only calibrated video cannot execute view selection or layout refinement, and the reported results are conditional on an input the method does not state. Please specify and evaluate the depth source, or explicitly scope the method and claims to RGB-D input.","section":"Sec. 2 and Sec. 3, Eq. (1), Eq. (7), Eq. (13)"},{"comment":"The implementation details in Appendix B set M=3, Tmid=15, and the layout loss coefficients; Sec. 4.2 states that the fusion-step ablation was conducted on Replica. The main Replica results in Tables 1-2 are on the same dataset. If the same scenes were used to choose Tmid and the loss weights, the reported improvements are partially a result of tuning on the evaluation set. Please report a separate validation split or show that the results are stable across a range of hyperparameters.","section":"Sec. 4.2 and Appendix B (hyperparameters)"},{"comment":"No standard deviations or number of seeds are reported for any experiment. The generation process is stochastic (flow matching with random initialization), and several reported differences are small (e.g., CD 39.37 vs 40.98 in Table 1, NC 75.17 vs 74.86). Please report at least three seeds with mean and std, and state whether the differences are significant.","section":"All tables"}],"minor_comments":[{"comment":"The two lists of loss coefficients are inconsistent. The first list gives λb=2.0 and λp=0.04; the second gives λb=0.04 and λp=0.5. Eq. (13) also includes L_reg with no coefficient. Please unify.","section":"Appendix B"},{"comment":"If the sum of signs in Eq. (10) is zero, Iunobs(u) may be empty and the average in Eq. (11) is undefined. Add a tie-breaking rule (e.g., fall back to the anchor or to uniform averaging).","section":"Sec. 3.2, Eq. (10)-(11)"},{"comment":"'Based on the selected views, its performs step-efficient...' should read 'it performs.'","section":"Fig. 2 caption"},{"comment":"Please specify whether the reported TFLOPs and time are for the full pipeline or only the flow model, and state the view-selection setting. The text says 'flow-model sampling' but the table header lacks units and scope.","section":"Table 4"},{"comment":"The sentence 'Compared with ICP, GS reduces CD by 43.8%/30.9%' is correct as an ablation, but it conflicts with the abstract's wording. Please harmonize the claim.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is well structured and the selection-fusion-alignment decomposition is a useful contribution. However, the abstract's headline scene-level claim is mislabeled in the current tables, and the unstated depth dependence is serious enough that the method, as described, cannot be applied to RGB-only calibrated video. A careful revision that corrects the comparisons, specifies and evaluates a depth source, and adds error bars would make the paper acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent engineering paper. The selection–fusion–alignment pipeline on frozen SAM3D is new as a combination, and the ablations back the design choices. The scene-level numbers are real for what they actually compare, but the abstract's claim of \"43.8% / 30.9% over SAM3D with ICP alignment\" is not what Tables 2 and 6 show: those reductions are Ours-GS vs Ours-ICP under the same fusion. That is still a useful result — the Gaussian layout refinement beats ICP on top of their fusion — but it is not a comparison against SAM3D+ICP.\n\nThe deeper soft spot is depth. The problem definition lists observations as (I, M, C, K) with no depth, yet Eq. 1, Eq. 7, and Eq. 13 require per-pixel depth D_i. On Replica and ScanNet++ they read GT depth from the dataset. So the headline \"calibrated multi-view scene asset generation\" overreaches: for standard RGB-only video, you would need an external depth estimator, which is neither specified nor evaluated. This is a load-bearing gap, and it is fixable — state the depth source explicitly, or add a variant that runs an off-the-shelf monocular depth estimator.\n\nOther gaps are minor: no seeds or error bars, M, Tmid, and loss weights selected on Replica and reused on Replica, and code is promised but absent. The baseline set is fair — Trellis, Amodal3R, InstantMesh, ReconViaGen, and MD+SAM3D are the right comparisons. The instance-level gains over MD+SAM3D are modest (CD 40.98 to 39.37 on Replica) but consistent, and the efficiency saving near 20% under ten views is a real practical point.\n\nOverall, the method holds together as a pipeline, the components are well-motivated, and the ablations support the claims within their stated comparisons. The needed fixes are reporting and validation, not a broken derivation. I would take it if the authors correct the comparison wording in the abstract and either specify the depth source or evaluate performance with an external depth estimator.\n\nThis is worth a serious referee: the combination is new, the engineering is sound, and the gaps are addressable. A conditional accept after revision is the right call.","headline":"A solid training-free multi-view extension of SAM3D with a genuinely new selection–fusion–alignment combination and honest component ablations, but the abstract overstates the headline comparison and the method silently depends on depth it never specifies.","tokens_in":18219,"tokens_out":1763,"would_cite":true,"duration_ms":16438,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Scene-SAM3D turns a frozen single-view 3D generator into a multi-view scene generator without fine-tuning.","keywords":["multi-view 3D generation","scene asset generation","view selection","velocity fusion","Gaussian layout refinement","training-free","flow matching","calibrated cameras"],"falsifier":"Run Scene-SAM3D on a calibrated RGB-only capture without dataset-provided depth or masks and measure scene-level Chamfer distance against single-view and MultiDiffusion baselines; if the occupancy-based view selection and Gaussian layout refinement no longer help, the central claim fails. Alternatively, reproduce the scene-level comparison with the baseline stated in the abstract — SAM3D with ICP alignment using the same coverage-selected views — and check whether the 43.8%/30.9% reductions still hold when the reference is SAM3D+ICP rather than the pipeline's own ICP refinement.","tokens_in":17272,"feed_emoji":"🏠","tokens_out":5555,"duration_ms":46750,"temperature":0.7,"pith_summary":"The paper argues that calibrated multi-view scene asset generation can be solved by composing three training-free modules on top of a frozen single-view object generator: occupancy-coverage view selection, early-stop sign-consensus velocity fusion, and rigid-object Gaussian layout refinement. The central claim is that this selection–fusion–alignment decomposition lets a strong object prior (SAM3D) produce complete, layout-coherent indoor scenes despite occlusion, redundant observations, and cross-view inconsistencies. On two indoor datasets, the paper reports instance-level gains in Chamfer distance and completeness over single-view and MultiDiffusion baselines, scene-level CD reductions of 43.8% on Replica and 30.9% on ScanNet++ when the proposed Gaussian refinement replaces its own ICP alignment, and roughly 20% savings in flow-model FLOPs and wall-time. A sympathetic reader would care because the recipe is parameter-free: it upgrades existing feed-forward 3D generators to multi-view capture without retraining.","feed_headline":"Gaussian layout refinement cuts scene 3D error by up to 44%","feed_subtitle":"Training-free pipeline couples view selection, early-stop fusion, and rigid-object optimization — no fine-tuning.","key_machinery":"The load-bearing machinery is a set of three operators. (1) Occupancy-coverage view selection (Eqs. 1–3): masked depth points are voxelized in a shared world frame; the anchor maximizes merged occupied volume and each helper greedily adds the most uncovered voxels, so views are judged by 3D coverage rather than 2D mask area. (2) Step-efficient sign-consensus fusion (Eqs. 8–11): helper-view velocities are computed only before a midpoint Tmid, and within each latent cell only velocities whose sign matches the anchor (in anchor-observed cells) or a majority vote (elsewhere) are averaged, preventing incompatible canonical shapes from being mixed. (3) Rigid-object Gaussian layout refinement (Eqs.","core_discovery":"The central discovery is that multi-view evidence can be exploited in a frozen flow-matching generator without any training: pick one anchor view that maximizes world-space occupancy coverage, add helper views that add new voxels, fuse their velocity fields only in early sampling steps and only where per-cell signs agree with the anchor (or with a majority vote in anchor-unobserved regions), then leave geometry untouched and optimize only each object's rotation, scale, and translation in a differentiable Gaussian proxy against multi-view masks, RGB, depth, boundaries, and inter-object penalties. The paper claims this suppresses duplicated or layered structures from uniform velocity averaging","pith_inferences":["The same selection–fusion–alignment split is a general recipe: any flow- or diffusion-based single-view 3D generator could be wrapped in these three stages, so the contribution is not tied to SAM3D specifically.","Because the experiments read depth directly from the datasets, a natural next test is to substitute monocular depth estimation and run the pipeline on RGB-only footage; if the gains survive, the method becomes applicable to casual phone or robot captures.","The decoupling of shape from layout implies the Gaussian refinement stage could be reused as a post-process for any object-centric generator, independent of the fusion module that precedes it.","The sign-consensus rule suggests that spatially selective averaging of velocity fields (rather than uniform averaging) could improve other multi-condition latent-space generation tasks, such as text-guided or multi-image 3D generation."],"forward_implications":["A single-view 3D generator can produce complete scene assets from calibrated multi-view video without any fine-tuning, as long as depth and masks are available.","Occupancy-based view selection improves reconstruction even for single-view baselines, indicating that world-space 3D coverage is a better view-quality signal than 2D mask size.","Early-stop, sign-consensus fusion avoids the duplicated or layered geometry that uniform multi-view velocity averaging produces, while still recovering anchor-unobserved regions.","Refining only object-level similarity transforms (not geometry) yields large scene-layout gains — 43.8%/30.9% lower scene-level CD and 40.8%/42.5% higher IoU — and converges within 200 iterations.","Under ten-view inference, the step-efficient fusion reduces flow-model FLOPs and wall-time by nearly 20% versus exhaustive multi-view fusion."],"fun_headline_variants":["Training-free multi-view fusion cuts 3D scene error by 44%","No fine-tuning, just view selection and velocity fusion – 44% less error","Frozen generator + rigid optimization: 3D scene error down 44%","Scene-SAM3D: multi-view evidence slashes scene CD by 30–44%","Pick anchor views, fuse early, tune layout: 44% better scenes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The pipeline assumes a reliable per-pixel depth map D_i and an instance mask are available for every view used in view selection and layout refinement; the paper never specifies how depth is obtained for arbitrary calibrated RGB video, since on the two benchmark datasets it is read directly from the data.","fun_headline_variants_meta":{"raw":{"variants":["Training-free multi-view fusion cuts 3D scene error by 44%","No fine-tuning, just view selection and velocity fusion – 44% less error","Frozen generator + rigid optimization: 3D scene error down 44%","Scene-SAM3D: multi-view evidence slashes scene CD by 30–44%","Pick anchor views, fuse early, tune layout: 44% better scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1303,"prompt_tokens":748,"completion_tokens":555,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":458}},"tokens_in":492,"tokens_out":555,"duration_ms":5439,"temperature":1.0,"reasoning_tokens":458,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T19:52:38.467456+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Scene-SAM3D on a calibrated RGB-only capture without dataset-provided depth or masks and measure scene-level Chamfer distance against single-view and MultiDiffusion baselines; if the occupancy-based view selection and Gaussian layout refinement no longer help, the central claim fails. Alternatively, reproduce the scene-level comparison with the baseline stated in the abstract — SAM3D with ICP alignment using the same coverage-selected views — and check whether the 43.8%/30.9% reductions still hold when the reference is SAM3D+ICP rather than the pipeline's own ICP refinement.","supporting_citations":[],"review_version":1}