Pith. sign in

REVIEW 4 major objections 5 minor 32 references

GarmentGS: Point-Cloud Guided Gaussian Splatting for High-Fidelity Non-Watertight 3D Garment Reconstruction

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read By guiding Gaussian splats with a ten-minute point cloud, GarmentGS reconstructs single-layer, non-watertight garments ready for simulation.

desk verdict A plausible point-cloud-guided Gaussian splatting pipeline for non-watertight garment meshes, but the evaluation is too weak to back the headline numbers. read the letter →

arxiv 2505.02126 v2 pith:PPLWULPP submitted 2025-05-04 cs.CV

classification cs.CV
keywords 3DgarmentreconstructionGaussiansplattingmulti-viewstereodensepointcloudpriornon-watertightmeshsingle-layersurfacefabricsimulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [Section 4.2 / Table 1] 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.
  2. [Section 3.1 / Dense Point Cloud Reconstruction] 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.
  3. [Section 3.3] 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.
  4. [Section 4.1 / Table 1] 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.
minor comments (5)
  1. [Section 3.2 / Figure 2] 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.
  2. [Table 2] 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.
  3. [Section 4.1] 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.
  4. [Section 3.1 / Table 1] 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.
  5. [Abstract / Table 1] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the geometric prior and Gaussian optimization are both derived from the same multi-view images, but no output quantity is equivalent to an input by construction and the main metric is evaluated against independent ground-truth meshes.

full rationale

The derivation chain is: multi-view images → fast MVS dense point cloud → Gaussian movement/flattening/rotation (Eqs. 4–6) → mesh extraction → LOF trimming (Sec. 3.3). The dense point cloud is an intermediate representation produced from the same images that supervise the Gaussians, but this is standard self-contained reconstruction, not circularity: the point cloud is not defined in terms of the final mesh, and the final mesh is not used to define the point cloud. The load-bearing assumption that "points in the dense point cloud are strictly distributed on the surface" is a robustness assumption about prior quality; if the point cloud has holes or noise, the output can degrade, but that is a correctness/risk issue rather than a logical reduction of the result to its inputs. The Chamfer distance evaluation is computed against separate ground-truth meshes from DeepFashion3D, so the headline quantitative claim is externally benchmarked rather than fitted. The paper contains no load-bearing self-citations: references such as GaussianPro [4], 2DGS [9], and PGSR [3] are external prior works, and no uniqueness theorem or prior result by the same authors is invoked to force the method. Each component (movement, rotation, flattening, LOF) is ablated in Table 2, and the full system is compared against independent baselines. Thus no specific equation or claimed prediction reduces by construction to its own input, and the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method is largely a combination of existing techniques. The free parameters are few but hand-set, and the most consequential assumption is that the MVS point cloud is a faithful surface approximation. No new entities or physical objects are introduced.

free parameters (4)
  • alpha (loss weight for flattening) = 100
    Hand-set in Eq. 7 with no sensitivity analysis or justification beyond a fixed value.
  • beta (loss weight for normal alignment) = 0.1
    Hand-set in Eq. 7 with no sensitivity analysis; no ablation over alpha/beta is provided.
  • MVS simplification settings (depth map resolution, iteration counts) = unspecified
    Section 3.1 says resolution and iterations are reduced but gives no values; these directly control point cloud quality and the claimed 10-minute runtime.
  • LOF parameters (number of neighbors, contamination) = unspecified
    Section 3.3 uses LOF to detect outliers but does not specify the parameter values, which affect which faces are removed.
assumptions (4)
  • domain assumption The dense point cloud is an accurate and complete approximation of the garment surface, including at openings.
    Stated in Section 3.3 ('points in the dense point cloud are strictly distributed on the surface') and used in Eq. 4 and the LOF trimming; if false, the final mesh inherits point cloud errors.
  • domain assumption Flattening 3D Gaussians into 2D disks preserves sufficient rendering quality for the RGB loss.
    The method relies on the 2DGS approximation (Section 3.2, Eq. 5) without re-evaluating its validity for garment materials.
  • domain assumption The multi-view rendered images are consistent enough for the simplified MVS to produce a usable point cloud.
    Experiments use synthetic Blender renders with known lighting; real-world images may break the consistency assumption.
  • domain assumption A mesh can be extracted from the optimized Gaussians using an unnamed algorithm, and this mesh is suitable for LOF trimming.
    Section 3.3 refers to 'after mesh extraction' without naming or describing the algorithm, so the whole non-watertight single-layer claim depends on this unstated step.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GarmentGS: Point-Cloud Guided Gaussian Splatting for High-Fidelity Non-Watertight 3D Garment Reconstruction." pith.science (2026). https://pith.science/paper/PPLWULPP

@misc{pith2026250502126,
  author       = {Pith},
  title        = {Pith review of: GarmentGS: Point-Cloud Guided Gaussian Splatting for High-Fidelity Non-Watertight 3D Garment Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PPLWULPP}},
  note         = {Machine review of arXiv:2505.02126}
}
read the original abstract

Traditional 3D garment creation requires extensive manual operations, resulting in time and labor costs. Recently, 3D Gaussian Splatting has achieved breakthrough progress in 3D scene reconstruction and rendering, attracting widespread attention and opening new pathways for 3D garment reconstruction. However, due to the unstructured and irregular nature of Gaussian primitives, it is difficult to reconstruct high-fidelity, non-watertight 3D garments. In this paper, we present GarmentGS, a dense point cloud-guided method that can reconstruct high-fidelity garment surfaces with high geometric accuracy and generate non-watertight, single-layer meshes. Our method introduces a fast dense point cloud reconstruction module that can complete garment point cloud reconstruction in 10 minutes, compared to traditional methods that require several hours. Furthermore, we use dense point clouds to guide the movement, flattening, and rotation of Gaussian primitives, enabling better distribution on the garment surface to achieve superior rendering effects and geometric accuracy. Through numerical and visual comparisons, our method achieves fast training and real-time rendering while maintaining competitive quality.

Figures

Figures reproduced from arXiv: 2505.02126 by the authors.

Figure 1
Figure 1. Our method reconstructs 3D garments from multi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our method. We first (a) pull scattered 3D Gaussian primitives towards their nearest points in the dense [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison between ours and other methods on [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 14 canonical work pages

  1. [1]

    Jon Louis Bentley. 1975. Multidimensional binary search trees used for associative searching. Commun. ACM 18, 9 (1975), 509–517

  2. [2]

    Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and Jörg Sander. 2000. LOF: identifying density-based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data . 93–104

  3. [3]

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. 2024. PGSR: Planar- based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction. arXiv:2406.06521 [cs.CV] https://arxiv.org/abs/2406.06521

  4. [4]

    Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wen- ping Wang, and Xuejin Chen. 2024. GaussianPro: 3D Gaussian Splatting with Progressive Propagation. arXiv preprint arXiv:2402.14650 (2024)

  5. [5]

    CLO3D. 2024. CLO3D garment design software . https://www.clo3d.com

  6. [6]

    Silvano Galliani, Katrin Lasinger, and Konrad Schindler. 2015. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE international conference on computer vision . 873–881

  7. [7]

    Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. 2022. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems 35 (2022), 31841–31854

  8. [8]

    Kai He, Kaixin Yao, Qixuan Zhang, Jingyi Yu, Lingjie Liu, and Lan Xu. 2024. Dress- Code: Autoregressively Sewing and Generating Garments from Text Guidance. arXiv preprint arXiv:2401.16465 (2024)

Show all 32 references
  1. [9]

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao

  2. [10]

    Heewoo Jun and Alex Nichol. 2023. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)

  3. [11]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis

  4. [12]

    Taekyung Kim, Jaehoon Choi, Seokeon Choi, Dongki Jung, and Changick Kim

  5. [13]

    Lijuan Liu, Xiangyu Xu, Zhijie Lin, Jiabin Liang, and Shuicheng Yan. 2023. To- wards garment sewing pattern reconstruction from a single image. ACM Trans- actions on Graphics (TOG) 42, 6 (2023), 1–15

  6. [14]

    Zhen Liu, Yao Feng, Yuliang Xiu, Weiyang Liu, Liam Paull, Michael J Black, and Bernhard Schölkopf. 2023. Ghost on The Shell: An Expressive Representation of General 3D Shapes. arXiv preprint arXiv:2310.15168 (2023)

  7. [15]

    Maya. 2024. Autodesk Maya. https://www.autodesk.com/products/maya/

  8. [16]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Commun. ACM 65, 1 (2021), 99–106

  9. [17]

    Johannes L Schonberger and Jan-Michael Frahm. 2016. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition. 4104–4113

  10. [18]

    Johannes L Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys

  11. [19]

    Style3D. 2024. Style3D digital fashion solution . https://www.linctex.com

  12. [20]

    Kaixuan Wang and Shaojie Shen. 2018. Mvdepthnet: Real-time multiview depth estimation neural network. In 2018 International conference on 3d vision (3DV) . IEEE, 248–257

  13. [21]

    Qingshan Xu and Wenbing Tao. 2019. Multi-scale geometric consistency guided multi-view stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5483–5492

  14. [22]

    Ruigang Yang and Marc Pollefeys. 2003. Multi-resolution real-time stereo on commodity graphics hardware. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. , Vol. 1. IEEE, I–I

  15. [23]

    Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. 2018. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV) . 767–783

  16. [24]

    Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai

  17. [25]

    Zehao Yu, Torsten Sattler, and Andreas Geiger. 2024. Gaussian Opac- ity Fields: Efficient Adaptive Surface Reconstruction in Unbounded Scenes. arXiv:2404.10772 [cs.CV] https://arxiv.org/abs/2404.10772

  18. [26]

    Wenyuan Zhang, Yu-Shen Liu, and Zhizhong Han. 2024. Neural signed distance function inference through splatting 3d gaussians pulled on zero-level set. arXiv preprint arXiv:2410.14189 (2024)

  19. [27]

    Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. 2020. Deep Fashion3D: A Dataset and Benchmark for 3D Garment Reconstruction from Single Images. In Computer Vision – ECCV . doi:10.1007/978-3-030-58452-8_30

  20. [29]

    arXiv:2403.16964 [cs]

    GSDF: 3DGS Meets SDF for Improved Rendering and Reconstruction. arXiv:2403.16964 [cs]

  21. [2016]

    In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, Octo- ber 11-14, 2016, Proceedings, Part III 14

    Pixelwise view selection for unstructured multi-view stereo. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, Octo- ber 11-14, 2016, Proceedings, Part III 14 . Springer, 501–518

  22. [2021]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision

    Just a few points are all you need for multi-view stereo: A novel semi- supervised learning method for multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 6178–6186

  23. [2023]

    ACM Transactions on Graphics 42, 4 (Aug

    3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics 42, 4 (Aug. 2023), 1–14. doi:10.1145/3592433

  24. [2024]

    In Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers ’24

    2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers ’24. ACM, Denver CO USA, 1–11. doi:10.1145/3641519.3657428

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.