{"id":"07e961f9-8017-445d-bfa9-7eb450f58ab4","arxiv_id":"2505.02126","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GarmentGS combines a fast dense point cloud with 2D Gaussian splatting to reconstruct high-fidelity, non-watertight, single-layer garment meshes from multi-view images.","lead":"This paper describes a 3D garment reconstruction method that uses a fast multi-view stereo point cloud to pull and flatten Gaussian splatting primitives onto the clothing surface, then trims the extracted mesh into a single open fabric layer. The authors report roughly 11 minutes of total computation with higher rendering and geometry scores than existing Gaussian-based baselines on a 9-garment benchmark.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pipeline's single-layer, non-watertight garment claim depends entirely on the fast MVS point cloud being surface-accurate and complete; holes or noise at garment openings propagate through Eq.","rationale":"I agree with the reader's conditional verdict and with the identification of the dense point cloud prior as the weakest assumption. My stress-test pass sharpens this into a concrete, load-bearing failure mode: the point cloud is trusted without a recovery mechanism, and it is used both as the geometric target for Gaussian snapping and as the inlier reference for LOF. That makes the single-layer and non-watertight claims directly dependent on MVS completeness and accuracy, and the paper provides no evidence for those properties under realistic garment imaging. The evaluation on 9 synthetic Blender renders with rich texture, no error bars, and no point-cloud ablation cannot rule out this failure mode. The Gaussian-loss ablations in Table 2 are informative for the rendering-quality side of the claim but do not bound the point-cloud dependence. Therefore the verdict should remain CONDITIONAL: the idea is promising and the reported numbers are encouraging, but the central practical claim requires a stress test that the paper does not provide. I partially agree with the reader because they identify the point-cloud fidelity assumption but do not trace its specific amplification through Eq. 4 and LOF, nor propose the direct stress-test experiments that would settle it.","tokens_in":7485,"tokens_out":2194,"duration_ms":29463,"concrete_test":"Run GarmentGS on the same 9 garment models but render them with low-texture, nearly uniform fabric under flat lighting, plus a set of dark/black garments, to stress-test MVS patch matching. Measure (a) CD versus ground truth, (b) the fraction of ground-truth openings (neckline, sleeve, hem) that become closed, bridged, or distorted, and (c) the number of disconnected layers or inner fragments after LOF. A second analytic stress test: replace the MVS point cloud with a decimated ground-truth surface (10% of points) and with ground-truth plus 1% uniform noise, then run the Gaussian snapping and LOF steps; if either corruption breaks the single-layer non-watertight output or raises CD above 0.7, the method has insufficient margin for realistic MVS error.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is a direct pipeline: multi-view images -> fast MVS dense points -> Gaussian snapping (Eq. 4, Section 3.2) -> LOF mesh trimming (Section 3.3) -> non-watertight, single-layer, simulation-ready garment. The load-bearing condition is the assertion in Section 3.3 that \"points in the dense point cloud are strictly distributed on the surface.\" This condition is used twice: first, Eq. 4 snaps every Gaussian to its nearest MVS point, so any hole or noise in the point cloud pulls Gaussians to wrong locations; second, LOF uses that same point cloud as its sole inlier reference, so the trimmed mesh can only be as clean as the point cloud. Four failure modes route through this condition: (1) MVS holes in textureless or dark fabric regions will snap Gaussians into concavities or bridge openings, directly breaking the non-watertight wearability claim; (2) MVS noise just off the surface is treated as ground truth by LOF, with no independent photometric or mesh-consistency cleaning; (3) the fast MVS module is underspecified (no architecture, fusion rule, normal-estimation detail, or hole-filling step is given, and the 10-minute timing is not tied to a concrete baseline), so the claimed prior cannot be independently reproduced; (4) the evaluation uses only 9 synthetic Blender-rendered garments with rich texture and no error bars or per-scene numbers, so the reported CD=0.564 is measured precisely in the regime where the point-cloud prior works best. Table 2 ablates the Gaussian movement/rotation/flattening losses but never ablates the point cloud itself, so the method's margin against point-cloud error is unmeasured.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GarmentGS, a multi-view 3D garment reconstruction method built on 3D Gaussian Splatting. It first generates a dense point cloud with a fast, simplified MVS pipeline, then uses that point cloud to guide the Gaussian primitives: Eq. (4) snaps each Gaussian to its nearest point, Eq. (5) flattens Gaussians into 2D disks, and Eq. (6) aligns their normals to point-cloud normals. After mesh extraction, Section 3.3 applies LOF to remove outlier faces using the dense point cloud as the inlier reference, producing a non-watertight, single-layer mesh. Evaluations on 9 synthetic DeepFashion3D garments report PSNR 40.13, LPIPS 0.017, Chamfer distance 0.564, and qualitative single-layer/wearable/textured checkmarks, outperforming 3DGS, GOF, GSDF, 2DGS, and PGSR.","tokens_in":7798,"tokens_out":4368,"duration_ms":53571,"significance":"If the results hold, the contribution is practically relevant: a 10-minute dense point cloud prior, combined with Gaussian guidance and LOF trimming, could make 3DGS-based reconstruction directly usable for garment simulation and virtual try-on, which is a meaningful step beyond 2DGS and PGSR. The paper explicitly targets non-watertight, single-layer output, a distinctive and falsifiable claim, and the ablations in Table 2 isolate the effects of movement, rotation, and flattening. The high-level pipeline is clearly described. However, the significance is currently bounded by the lack of a controlled comparison against baselines receiving the same point-cloud prior, by the underspecification of the key MVS module, and by the qualitative nature of the central wearability and single-layer claims.","major_comments":[{"comment":"The headline comparisons are not controlled: GarmentGS receives a dense MVS point cloud computed from the same 100 views, while 3DGS, GOF, GSDF, 2DGS, and PGSR receive only images. The reported gains (PSNR 40.13 vs 36.33 for 2DGS; CD 0.564 vs 0.712) could therefore reflect the extra surface prior rather than the proposed Gaussian-optimization or LOF stages. Please add a baseline that is given the same point cloud (e.g., 2DGS or PGSR initialized with the MVS points) or ablate the point-cloud prior entirely. Table 2 currently varies only movement/rotation/flattening and never removes the point cloud, so it does not resolve this confound.","section":"Section 4.2 / Table 1"},{"comment":"The fast MVS module is the load-bearing first stage, but it is specified only as \"removing color information... reducing depth map resolution and decreasing optimization iteration counts,\" with no underlying algorithm name, architecture, depth-fusion strategy, normal-estimation method, or concrete hyperparameter values. The 10-minutes-versus-2-hours claim is not tied to a reproducible configuration. Please provide the full pipeline details or cite a specific baseline and state the exact modifications and settings; otherwise Eq. (4) and the LOF trimming cannot be independently reproduced or verified.","section":"Section 3.1 / Dense Point Cloud Reconstruction"},{"comment":"The statement that \"points in the dense point cloud are strictly distributed on the surface\" is an unsupported assumption. MVS point clouds from real garments typically have holes at openings, missing coverage on textureless fabric, and off-surface noise. Because LOF uses this same point cloud as the only inlier reference, any such error directly propagates into the final mesh, potentially trimming valid fabric at necklines or sleeves or leaving internal fragments. Please report point-cloud accuracy and completeness against the ground-truth meshes, or demonstrate robustness to injected noise and holes, before claiming \"simulation-ready\" output.","section":"Section 3.3"},{"comment":"The quantitative evaluation is limited to 9 synthetic Blender-rendered garments and reports only aggregate scores, with no per-scene breakdown, error bars, or statistical significance. Synthetic renders with rich textures are the most favorable regime for the MVS prior, so the claimed superiority over 2DGS and PGSR is not yet established for real capture conditions. Additionally, the \"Single-layer,\" \"Wearable,\" and \"Texture\" columns are filled by unspecified criteria. Please define these metrics objectively (e.g., number of connected components, open-boundary count, Euler characteristic, UV-map existence), include per-garment results with variance, and either add a real-image sequence or clearly state the synthetic-only scope.","section":"Section 4.1 / Table 1"}],"minor_comments":[{"comment":"The loss names appear as Ł in the equations but as L in the text, and Eq. (5) uses a 1-norm while Figure 2 appears to show a 2-norm-style expression. Please unify the notation and ensure all symbols (m_i, n_i, s_i) are defined in the figure caption.","section":"Section 3.2 / Figure 2"},{"comment":"The row semantics are ambiguous: rows labeled \"Movement,\" \"Rotation,\" and \"Flattening\" could mean \"only this component\" or \"without this component,\" and the presence of both \"Movement\" and \"No-Movement\" rows makes the intended interpretation unclear. Please clarify in the caption or text.","section":"Table 2"},{"comment":"The dataset is called DeepFashion3D-v2 but is cited to the original DeepFashion3D paper [27]. If a v2 version exists, please provide the correct citation or explain the relationship between the v2 data and the original dataset.","section":"Section 4.1"},{"comment":"The time reporting is ambiguous: Section 3.1 says point-cloud reconstruction takes about 10 minutes, but Table 1 reports a total time of 11.3m for Ours. Please clarify whether the 11.3m includes MVS, Gaussian training, and mesh extraction, and state what exactly the 2-hour \"traditional method\" [18] corresponds to.","section":"Section 3.1 / Table 1"},{"comment":"The abstract claims \"fast training,\" but Table 1 shows 11.3m for Ours versus 3.3m for 3DGS and 3.9m for 2DGS. Please clarify whether \"fast\" is meant relative to NeRF/SDF-based methods rather than to other 3DGS baselines.","section":"Abstract / Table 1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is simple and potentially useful for the garment-reconstruction community, and the paper is within scope for ICMR. However, the experimental protocol needs a controlled comparison where the baselines also receive the dense point-cloud prior, the MVS stage needs enough detail to be reproducible, and the central single-layer/wearable claims need objective quantification. The paper would also benefit from code release, though I do not treat that as a hard requirement. Given these load-bearing issues, I recommend major revision rather than rejection, because the approach is defensible and the missing evidence appears obtainable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nYou asked about GarmentGS. The short version: it's a sensible combination of known pieces—GaussianPro-style snapping, 2DGS flattening, plus an LOF cleanup—driven by a fast MVS point cloud, aimed at generating non-watertight, single-layer garment meshes. That is genuinely new as a pipeline, even if none of the components are. The 10-minute point cloud reconstruction claim is interesting, and the idea of using the point cloud to both position Gaussians and trim internal mesh fragments is coherent.\n\nThe problems are mostly in the evidence. The MVS module is underspecified: no architecture, no fusion rule, no hole-filling, and the 10-minute figure isn't tied to a concrete baseline. The comparison in Table 1 is apples-to-oranges because the proposed method gets a dense point cloud prior that none of the baselines receive. Even with that advantage, only 9 synthetic Blender-rendered garments are reported, with no error bars or per-scene breakdown. The ablations in Table 2 drop each loss term but never remove the point cloud itself, so we never see how much margin actually comes from that prior. Section 3.3 asserts that points in the dense point cloud are 'strictly distributed on the surface,' but nothing demonstrates that; if that claim is wrong, the LOF trimming will cut valid fabric or leave inner fragments.\n\nThese are real soft spots, but not showstoppers in principle. For a five-page short paper, the missing details are partially forgivable. The pipeline is clear, the figures look plausible, and the non-watertight single-layer output is a useful goal. The problem is that the headline numbers (PSNR 40 vs 36, CD 0.564) are not trustworthy as reported because of the unfair comparison and the lack of variance measures.\n\nI'd send this to peer review—the approach deserves a serious look—but I'd ask for a major revision: more scenes (including real captures), error bars, an ablation without the point cloud, and at least one baseline given the same prior. As is, it reads like a solid workshop paper rather than a fully backed conference contribution.\n\nBest","headline":"A plausible point-cloud-guided Gaussian splatting pipeline for non-watertight garment meshes, but the evaluation is too weak to back the headline numbers.","tokens_in":8397,"tokens_out":2453,"would_cite":false,"duration_ms":29329,"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":"By guiding Gaussian splats with a ten-minute point cloud, GarmentGS reconstructs single-layer, non-watertight garments ready for simulation.","keywords":["3D garment reconstruction","Gaussian splatting","multi-view stereo","dense point cloud prior","non-watertight mesh","single-layer mesh","surface reconstruction","fabric simulation"],"falsifier":"Render one garment with an open neckline or loose sleeve, run the full pipeline, and inspect whether the mesh opening is preserved: if the point cloud misses coverage at the opening, nearby Gaussians are snapped across the gap and the LOF trim removes valid surface, so the output becomes closed or double-layered rather than non-watertight and single-layer.","tokens_in":7258,"feed_emoji":"👗","tokens_out":5623,"duration_ms":58197,"temperature":0.7,"pith_summary":"GarmentGS claims that a cheap, incomplete surface prior—a dense point cloud computed in about ten minutes from multi-view images—is enough to turn 3D Gaussian splatting from a rendering method into a garment reconstruction method. The paper's core assertion is that moving, flattening, and rotating Gaussian primitives to match the point cloud produces a high-fidelity, non-watertight, single-layer mesh that can go straight into fabric simulation, while avoiding the several-hour cost of traditional multi-view stereo and the roughly one-hour cost of SDF-based alternatives. A sympathetic reader would care because current automated garment reconstruction either produces multi-layered or watertight meshes that need manual cleanup, or is too slow for practical use. If the claim holds, the pipeline makes simulation-ready garments from ordinary multi-view photos in minutes.","feed_headline":"One point cloud steers Gaussians into wearable single-layer garments","feed_subtitle":"A 10-minute point cloud plus Gaussian splatting yields non-watertight meshes ready for fabric simulation.","key_machinery":"The load-bearing object is the dense point cloud used as a surface prior, combined with three manipulations of Gaussian primitives: movement (snap each Gaussian's mean to the nearest point in the point cloud via a KD-tree), flattening (compress the smallest scale $s_i$ of the covariance matrix through $\\mathcal{L}_{\\text{thin}} = \\|\\min(s_1,s_2,s_3)\\|_1$), and rotation (align Gaussian disk normal $\\mathbf{n}_i$ with point normal $\\mathbf{m}_i$ through $\\mathcal{L}_{\\text{normal}} = 1 - |\\mathbf{m}_i \\cdot \\mathbf{n}_i|$). The total loss is $\\mathcal{L} = \\mathcal{L}_{\\text{RGB}} + \\alpha \\mathcal{L}_{\\text{thin}} + \\beta \\mathcal{L}_{\\text{normal}}$ with $\\alpha=100$, $\\beta=0.1$. After meshing, LOF uses the point cloud as the inlier reference to delete outlier faces.","core_discovery":"The central claim is that dense point clouds, obtained in about ten minutes from multi-view garment images, can be used as a geometric prior that makes Gaussian splatting reconstruct a garment surface rather than a volumetric cloud: each Gaussian is moved to its nearest point in the point cloud (Eq. 4), flattened into a 2D disk, and rotated so its normal matches the point normal; after mesh extraction, a local-outlier factor filter removes internal fragmented faces, yielding a single-layer, non-watertight mesh. The paper reports this outperforming recent Gaussian-based surface methods in PSNR, LPIPS, and Chamfer distance while being the only method marked single-layer, wearable, and textured.","pith_inferences":["The same point-cloud-guided movement-flattening-rotation recipe should transfer to other thin-shell objects such as hats, gloves, or shoes, where non-watertight single-layer output is also the desired end product.","The method's accuracy should degrade most at garment openings and thin protrusions, because those are exactly where a coarse point cloud has the least coverage; this is testable by comparing reconstruction quality against view density around the neckline.","Since the point cloud is the only geometric constraint, replacing the simplified multi-view stereo module with a learned single-view depth estimator is a natural extension, and the speed gain would be even larger if the estimator runs faster than ten minutes."],"forward_implications":["Garment meshes come out single-layer and non-watertight, so they can pass directly to fabric simulation and virtual try-on without manual cleanup.","Total pipeline time is about eleven minutes of training plus ten minutes of point-cloud reconstruction, well under the roughly one hour reported for SDF-based Gaussian methods.","The method is the only one in the comparison that simultaneously yields a textured, wearable, single-layer mesh.","Geometric accuracy improves because Gaussians are pulled onto the surface instead of floating in the volume, as shown by the lower Chamfer distance in the paper's table."],"supporting_citations":[{"why":"Base 3D Gaussian Splatting representation, rendering loss, and adaptive density control that the method builds on.","marker":"[11]"},{"why":"Source of the idea to use dense point clouds to adjust the position and rotation of Gaussian primitives.","marker":"[4]"},{"why":"Supplies the 2D Gaussian flattening strategy and the thinning regularization that turns ellipsoids into disks.","marker":"[9]"},{"why":"Provides the local outlier factor algorithm used to remove internal fragmented faces after mesh extraction.","marker":"[2]"},{"why":"Traditional multi-view stereo baseline reported to take about two hours, the speed comparison for the ten-minute point-cloud module.","marker":"[18]"},{"why":"DeepFashion3D dataset used for quantitative evaluation of rendering and geometric accuracy.","marker":"[27]"},{"why":"KD-tree structure used for the nearest-point queries that snap each Gaussian to the point cloud.","marker":"[1]"}],"fun_headline_variants":["10-minute point cloud guides Gaussians to single-layer garments","Point-cloud-guided Gaussian splatting for non-watertight 3D garments","Fast point cloud steers Gaussians into wearable garment surfaces","Gaussian splatting meets point clouds for rapid garment reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the dense point cloud generated in roughly ten minutes is accurate and complete enough to represent the garment surface, because every Gaussian is pinned to its nearest point and the final mesh-trimming step treats the point cloud as the trusted surface.","fun_headline_variants_meta":{"raw":{"variants":["10-minute point cloud guides Gaussians to single-layer garments","Point-cloud-guided Gaussian splatting for non-watertight 3D garments","Fast point cloud steers Gaussians into wearable garment surfaces","Gaussian splatting meets point clouds for rapid garment reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1455,"prompt_tokens":878,"completion_tokens":577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":505}},"tokens_in":494,"tokens_out":577,"duration_ms":6366,"temperature":1.0,"reasoning_tokens":505,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T01:01:56.514159+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render one garment with an open neckline or loose sleeve, run the full pipeline, and inspect whether the mesh opening is preserved: if the point cloud misses coverage at the opening, nearby Gaussians are snapped across the gap and the LOF trim removes valid surface, so the output becomes closed or double-layered rather than non-watertight and single-layer.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DeepFashion3D dataset used for quantitative evaluation of rendering and geometric accuracy."}],"review_version":1}