Pith. sign in

REVIEW 4 major objections 5 minor 27 references

A feedforward dual-branch generator, trained on cross-pose garment videos and topology-edited images, transfers textures across non-isometric image–mesh pairs and bakes them into production-ready PBR maps, reporting the best KID and FID amo

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 20:39 UTC pith:5LLHIN64

load-bearing objection A genuinely new data recipe for garment texture transfer across pose/topology mismatch; the central claim mostly holds up, but the cross-topology mechanism leans on an unvalidated black-box editor and the quantitative case is thin. the 4 major comments →

arxiv 2511.18765 v3 pith:5LLHIN64 submitted 2025-11-24 cs.CV cs.AI

NI-Tex: Non-isometric Image-based Garment Texture Generation

Hui Shan , Ming Li , Haitao Yang , Kai Zheng , Sizhe Zheng , Yanwei Fu , Xiangru Huang This is my paper
classification cs.CV cs.AI
keywords non-isometric texture generationgarment texturePBR material generationcross-topology augmentationuncertainty-guided bakingview selectionimage-conditioned texture generation3D garment videos
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that garment-texture generation no longer requires the input image and target mesh to be similar in pose, topology, or geometry. NI-Tex, a feedforward diffusion-based generator, is fine-tuned on cross-pose garment video frames and on images whose garment shape has been edited by a commercial image editor while the texture is preserved. The claim, supported by KID and FID comparisons against several commercial and published baselines, is that the model transfers texture identity faithfully across non-isometric pairs and fuses multi-view predictions into seamless albedo, roughness, and metallic maps. If true, industrial 3D garment meshes—which already cover most real-world clothing geometries but lack texture variety—could be re-textured directly from arbitrary photos.

Core claim

The central claim is that non-isometric image-based garment texture generation can be solved by constructing the right training data rather than by aligning geometry. NI-Tex learns invariance from two sources: pairs of frames of the same garment under different human poses, drawn from physically simulated 3D Garment Videos built on BEDLAM, and images whose garment topology has been synthetically altered by an editing model while the texture identity is kept. The trained dual-branch U-Net takes the input image as a guidance branch and multi-view normal and position maps as geometric conditions, outputting albedo, roughness, and metallic maps. A separate uncertainty-quantification network, tra

What carries the argument

The load-bearing mechanism is the training-pair construction: cross-frame sampling from physically simulated garment videos supplies pose and deformation invariance with consistent albedo ground truth, while cross-topology augmentation via image editing supplies topology invariance. The network is a dual-branch U-Net with Multi-Channel Aligned Attention (MCAA), an attention module that injects reference texture features into the albedo channel and lets the metallic-roughness channel share the same alignment. A switchable MR channel lets the model train on albedo alone when edited images have inconsistent reflections. The iterative baking loop uses a per-pixel uncertainty predictor—trained by

Load-bearing premise

The whole non-isometric capability rests on the assumption that the commercial image editor used for augmentation can change a garment's shape while faithfully preserving its texture identity—and the paper's own editing principles concede that the editor sometimes produces texture swaps, layering confusion, or extraneous body parts.

What would settle it

Collect a test set where the image prompt and target mesh share no topology and the texture is a precisely repeated logo, run NI-Tex, and inspect whether the logo bends, duplicates, or drifts at seams; if the model consistently redraws the logo instead of transferring it, the cross-topology gain would disappear in an ablation that removes the edited training images.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Industrial garment meshes with limited texture variety can be re-textured from arbitrary input photos without mesh deformation or topology matching.
  • Cross-pose texture consistency is learnable from simulated garment video sequences even when roughness and metallic values are assigned as globally uniform random values.
  • Uncertainty-guided iterative baking produces more complete texture maps than coverage-based view selection, reducing holes and blurriness from multi-view fusion.
  • The approach works on wild images, such as masked DeepFashion2 photos, and on meshes generated by text-to-3D models, not only on clean rendered views.
  • The switchable albedo-only training mode avoids distilling unreliable metallic and roughness attributes from edited images that lack consistent surface reflections.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reliance on a commercial image editor means the cross-topology capability is bounded by that editor's texture-preservation quality; outside the curated examples, the model may learn to redraw textures rather than transfer them.
  • The same recipe—paired frames plus shape-preserving image edits—could be transferred to other object categories, such as shoes or bags, where small topology changes matter and paired video data or simulation is available.
  • Because roughness and metallic values in the garment videos are globally uniform and randomly assigned, the model's MR generalization is likely weaker than its albedo generalization; the cross-mixing with Objaverse and TexVerse data is only a partial remedy.
  • The contribution of each pipeline stage could be isolated by ablating cross-pose video pairs and cross-topology edits separately, quantifying how much of the KID/FID gain comes from each.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes NI-Tex, a feedforward dual-branch method for image-conditioned PBR texture generation on garment meshes under non-isometric image–mesh pairs. It builds a physically simulated '3D Garment Videos' dataset from BEDLAM for cross-pose supervision, uses Nano Banana-edited images for cross-topology augmentation, and introduces an uncertainty-aware iterative baking procedure with view selection and reweighting. Experiments compare against Paint3D, Hyper3D, Hunyuan3D, and Meshy, reporting the best KID and FID in Table 1 and showing qualitative improvements on industrial and generated meshes.

Significance. If the claims hold, the paper addresses a real gap: existing image-to-texture pipelines degrade under pose or topology mismatch, and NI-Tex is a plausible feedforward solution. The physically simulated garment-video data with consistent albedo across frames is a reasonable external grounding for cross-pose learning, and the uncertainty-guided iterative baking is an interesting contribution that goes beyond simple coverage-based view selection. The paper is reproducible-sounding in architecture and data, and the authors state code and dataset release plans. However, the central non-isometric claim currently rests on a black-box image editor with no quantitative validation, and the headline quantitative evidence is a small-scale table without error bars.

major comments (4)
  1. [Section 4.1 (Cross-topology Augmentation)] The entire cross-topology capability is built on the assumption that Nano Banana can change garment topology while preserving texture identity. This is load-bearing: no other component in the pipeline addresses topology mismatch, and the paper provides no citation, no validation, no failure-rate statistics, and no filtering criterion for the 50K edited images. The paper's own design principles admit that the editor can cause texture swaps, inner/outer confusion, and extraneous body parts. If the editor's texture preservation is inconsistent outside curated examples, the training signal becomes 'redraw the texture from a corrupted image' rather than 'transfer the texture,' and the claimed non-isometric robustness is editor-dependent. I ask for a quantitative evaluation of editor fidelity (e.g., LPIPS/CLIP similarity between input and edited garment regions on a held-out set), a failure-ra
  2. [Section 5.3, Table 1] The quantitative claim 'significantly outperforms existing baselines' is not supported by Table 1 as reported. Only 10 image prompts/objects are used, with no standard deviations, confidence intervals, or significance tests. The KID margin over Meshy is 0.0364 vs 0.0383, which is small relative to typical KID variability. Moreover, it is unclear whether the 10 test cases specifically isolate non-isometric image–mesh pairs; a mixed set of 10 objects does not enable a claim about cross-topology robustness. I ask for error bars across multiple evaluation runs, a breakdown by topology-mismatch severity, and a larger test set (or a user study) to substantiate the central claim.
  3. [Section 4.1, Eq. (3), Appendix A.1 (MR Rectification)] The MR supervision is assigned as random per-frame uniform roughness with metallic=0, and then rectified by sampling a single foreground pixel from the condition frame. This is not ground-truth material information; it is an ad hoc proxy. The paper itself notes that the model tends to overfit these constant MR values and must be supplemented with Objaverse/TexVerse data. Given the paper's claim of generating production-ready PBR materials, the roughness/metallic quality needs evidence beyond qualitative images. Please provide an ablation or quantitative analysis showing that the learned MR maps are not simply predicting the uniform prior or a constant value.
  4. [Section 5.2, Figure 10 / Appendix B.4] The baking-strategy evaluation is limited. The quantitative comparison reports PSNR on a single 'worst viewpoint' per example, and the greedy metrics i_UQ and i_cvg in Eq. (7) score candidate views, not the final baked texture. There is no sensitivity analysis for c_j in Eq. (9), and the stopping thresholds epsilon and N_view are deferred to an appendix that is not provided in the main text. Since the iterative baking is one of the three main contributions, its benefit should be evaluated on the whole mesh, not only on one selected test view, and with multiple seeds to account for greedy selection randomness.
minor comments (5)
  1. [General] Nano Banana is used throughout as a black-box external tool but is never cited or described technically. At minimum, a reference or version/API identifier is needed for reproducibility.
  2. [Section 4.2] The sentence 'we train an Uncertainty Quantification (UQ) model' and the surrounding text contain typographical and grammatical issues (e.g., 'our train an'), which should be corrected. The description of the latent-code optimization in Eq. (4) also lacks details on initialization, optimizer, and stopping criterion.
  3. [Section 5.1] The training setup is described as 8 H200 GPUs for 10 days with batch size 2, but no learning rate, noise schedule, or fine-tuning details are given. These details are needed for reproducibility beyond a high-level description.
  4. [Section 4.1] The claim that cross-frame pairs expand to 'tens of billions of training samples' is combinatorial and not tied to the actual training budget. Please clarify how these pairs are sampled or whether this is simply an upper bound.
  5. [Section 5.3] The description 'using a fixed set of 42 random seeds' is ambiguous. It is not clear whether the 42 seeds are used for inference, for KID estimation, or for bootstrapping confidence intervals. Please clarify.

Circularity Check

0 steps flagged

No significant circularity: the training signals and evaluations are externally grounded.

full rationale

The paper's derivation chain is not circular. (1) Cross-pose training samples condition and supervision frames from BEDLAM sequences with a shared albedo texture and external PBR supervision; no predicted quantity is reinserted as its own training target. (2) Cross-topology augmentation uses Nano Banana to edit condition images while supervision remains the original PBR maps — an external (though uncited and unvalidated) source of topology variation; the trained model is later evaluated on independent DeepFashion2 prompts and industrial/Hunyuan meshes, so the claimed robustness does not reduce to the augmentation itself. (3) The UQ model is the closest self-referential component: its training errors are simulated with the same texture generator it later reweights. But the supervision for UQ is per-pixel SSIM against ground-truth maps, an external target, and the baking comparison is evaluated with PSNR on selected test views, not with UQ's own outputs. Thus this is self-training/calibration, not a definitional or fitted-input circle. (4) The cited references are external (BEDLAM, Objaverse, TexVerse, Hunyuan3D, MaterialMVP, etc.); no load-bearing result is justified only by a self-citation. The unvalidated Nano Banana fidelity assumption is a robustness/validity risk, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

Genuine external grounding for the central claim: BEDLAM, Objaverse, and TexVerse data plus Nano Banana-edited prompts supervised by original textures. But the PBR supervision is constructed rather than measured — roughness drawn from U(0,1) per frame, metallic fixed to 0, then MR rectification collapses supervision to one uniform value. The baking stage adds hand-set view weights (1, 0.5, 0.25, 0.125, 0.1) and an unspecified stopping threshold with no sensitivity analysis. No new physical entities are introduced; the UQ model and 3D Garment Videos are training artifacts, not explanatory postulates.

free parameters (6)
  • per-frame roughness value = U(0,1) random draw per frame (Eq. 3)
    Garment 'ground truth' roughness in 3D Garment Videos is a uniform random sample per frame, not a measured material property; the model is trained to reproduce a random constant.
  • metallic value = 0 for all garments (Eq. 3)
    All garments are assumed non-metallic; this removes metallic variation from the garment training signal.
  • MR rectification reference value = single foreground pixel sampled from condition frame MR (App. A.1)
    The supervision MR for the target frame is replaced by one sampled value; cross-frame MR supervision is thereby forced to a uniform constant.
  • loss balancing factor alpha = 2 (Eq. 6)
    Set to smooth the loss curve when alternating L1/L2; no sensitivity analysis is given.
  • view weighting scores c_j = 1, 0.5, 0.25, 0.125, 0.1 (Eq. 9)
    Hand-chosen attenuation of non-frontal/back views during baking; no ablation is reported.
  • iterative baking stop threshold epsilon and max views N_view = not specified
    Stopping criteria are mentioned ('More details are deferred to Appendix') but no values or sensitivity analysis appear in the provided appendix text.
axioms (5)
  • domain assumption BEDLAM's physically simulated garment deformations cover the pose/geometry distribution needed for garment texture transfer.
    Sec. 4.1: the entire cross-pose augmentation rests on simulation data being a valid proxy for real garment deformation; the paper itself states generalization to complex rigid deformations is limited (Sec. 6).
  • domain assumption Nano Banana edits preserve texture identity while altering garment topology.
    Sec. 4.1 lists three heuristics to 'avoid incorrect distillation'; the editor's texture preservation is assumed, not verified quantitatively.
  • ad hoc to paper A uniform scalar per garment is an adequate roughness/metallic supervision signal.
    Eq. 3 assigns roughness ~ U(0,1) and metallic = 0; MR rectification (App. A.1) then replaces all foreground pixels with a single sampled value. The 'versatile PBR materials' claim inherits this assumption.
  • domain assumption Per-pixel SSIM between generated and ground-truth texture maps is a valid per-pixel uncertainty label.
    App. A.2, Eq. 8: UQ supervision is the per-pixel SSIM map; artifacts like holes or blur are assumed to track SSIM.
  • standard math Stable Diffusion 2.1 denoising objective and Hunyuan3D/MaterialMVP MCAA cross-attention architecture transfer to the garment texture domain.
    Sec. 3.2, Eqs. 1-2, 5-6: the backbone and attention module are adopted as-is from prior work.

pith-pipeline@v1.3.0-alltime-deepseek · 11555 in / 25004 out tokens · 237549 ms · 2026-08-03T20:39:44.898113+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of NI-Tex: Non-isometric Image-based Garment Texture Generation." pith.science (2026). https://pith.science/paper/5LLHIN64

@misc{pith2026251118765,
  author       = {Pith},
  title        = {Pith review of: NI-Tex: Non-isometric Image-based Garment Texture Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5LLHIN64}},
  note         = {Machine review of arXiv:2511.18765}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing industrial 3D garment meshes already cover most real-world clothing geometries, yet their texture diversity remains limited. To acquire more realistic textures, generative methods are often used to extract Physically-based Rendering (PBR) textures and materials from large collections of wild images and project them back onto garment meshes. However, most image-conditioned texture generation approaches require strict topological consistency between the input image and the input 3D mesh, or rely on accurate mesh deformation to match to the image poses, which significantly constrains the texture generation quality and flexibility. To address the challenging problem of non-isometric image-based garment texture generation, we construct 3D Garment Videos, a physically simulated, garment-centric dataset that provides consistent geometry and material supervision across diverse deformations, enabling robust cross-pose texture learning. We further employ Nano Banana for high-quality non-isometric image editing, achieving reliable cross-topology texture generation between non-isometric image-geometry pairs. Finally, we propose an iterative baking method via uncertainty-guided view selection and reweighting that fuses multi-view predictions into seamless, production-ready PBR textures. Through extensive experiments, we demonstrate that our feedforward dual-branch architecture generates versatile and spatially aligned PBR materials suitable for industry-level 3D garment design.

Figures

Figures reproduced from arXiv: 2511.18765 by Haitao Yang, Hui Shan, Kai Zheng, Ming Li, Sizhe Zheng, Xiangru Huang, Yanwei Fu.

Figure 1
Figure 1. Figure 1: NI-Tex takes an image prompt and a target garment geometry as input, and generates high-quality PBR textures [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Texture generation becomes unreliable when the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of NI-Tex. In the data space (top), we construct our non-isometric training dataset from 3D Garment Videos by randomly selecting two frames, one as the condition 3D frame and the other as the supervision 3D frame, to enhance the model’s generalization across human poses, geometric deformations, and lighting variations. To further improve robustness to different garment topology, we apply Nano Bana… view at source ↗
Figure 5
Figure 5. Figure 5: We render each garment mesh from six view [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: We use Nano Banana to edit garment topol [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: (left) In the model training pipeline, we use an illuminated image from the condition 3D frame as the reference image, which is encoded by the guidance branch to perform texture transfer to the training branch. The supervision 3D frame provides multi-view normal and position images as geometric constraints, and its PBR (albedo and MR) maps for supervision. During supervision, MR Rectification is applied to… view at source ↗
Figure 7
Figure 7. Figure 7: We use 4D-Dress as our test set to evaluate the [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: (left) NI-Tex results using industrial well-rendered images (top) and DeepFashion2 wild images masked by SAM2 (bottom) as prompts, with industrial meshes as targets. NI-Tex produces realistic industry-level PBR materials and aligns textures across complex garment surfaces. The top also shows robust cross-topology generation under consistent poses. (right) NI-Tex results using DeepFashion2 wild images as pr… view at source ↗
Figure 9
Figure 9. Figure 9: We randomly select one MR image from the con [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: We compare the 10 views selected by the coverage-based strategy with those selected by our strategy. We then [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Texture generation results on industrial meshes using wild images from DeepFashion2 as image prompts. NI [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Texture generation results on Hunyuan-generated meshes using wild images from DeepFashion2 as image [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Multi-view visualization for industrial meshes (using well-render image prompts). [PITH_FULL_IMAGE:figures/full_fig_p015_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Multi-view visualization for industrial meshes (using image prompts from DeepFashion2). [PITH_FULL_IMAGE:figures/full_fig_p016_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Multi-view visualization for Hunyuan-generated meshes (using image prompts from DeepFashion2). [PITH_FULL_IMAGE:figures/full_fig_p017_15.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 8 linked inside Pith

  1. [1]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13142–13153, 2023

    2023

  2. [2]

    Texverse: A universe of 3d objects with high-resolution textures.arXiv preprint arXiv:2508.10868, 2025

    Yibo Zhang, Li Zhang, Rui Ma, and Nan Cao. Texverse: A universe of 3d objects with high-resolution textures.arXiv preprint arXiv:2508.10868, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Hunyuan3d 2.1: From images to high-fidelity 3d assets with production-ready pbr material

    Team Hunyuan3D, Shuhui Yang, Mingxin Yang, Yifei Feng, Xin Huang, Sheng Zhang, Zebin He, Di Luo, Haolin Liu, Yunfei Zhao, et al. Hunyuan3d 2.1: From images to high-fidelity 3d assets with production-ready pbr material. arXiv preprint arXiv:2506.15442, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Bedlam: A synthetic dataset of bodies exhibit- ing detailed lifelike animated motion

    Michael J Black, Priyanka Patel, Joachim Tesch, and Jin- long Yang. Bedlam: A synthetic dataset of bodies exhibit- ing detailed lifelike animated motion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726–8737, 2023

    2023

  5. [5]

    Texture: Text-guided texturing of 3d shapes

    Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. InACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023

    2023

  6. [6]

    Text2tex: Text- driven texture synthesis via diffusion models

    Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text- driven texture synthesis via diffusion models. InProceed- ings of the IEEE/CVF international conference on com- puter vision, pages 18558–18568, 2023

    2023

  7. [7]

    Easi-tex: edge-aware mesh texturing from single-image

    Sai Raj Kishore Perla. Easi-tex: edge-aware mesh texturing from single-image. 2024

    2024

  8. [8]

    Paint3d: Paint anything 3d with lighting-less texture diffu- sion models

    Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffu- sion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4252– 4262, 2024

    2024

  9. [9]

    Fan- tasia3d: Disentangling geometry and appearance for high- quality text-to-3d content creation

    Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fan- tasia3d: Disentangling geometry and appearance for high- quality text-to-3d content creation. InProceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023

    2023

  10. [10]

    Matlaber: Material-aware text-to-3d via latent brdf auto- encoder.arXiv preprint arXiv:2308.09278, 2023

    Xudong Xu, Zhaoyang Lyu, Xingang Pan, and Bo Dai. Matlaber: Material-aware text-to-3d via latent brdf auto- encoder.arXiv preprint arXiv:2308.09278, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering

    Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. InProceedings of the ieee/cvf conference on computer vi- sion and pattern recognition, pages 4347–4356, 2024

    2024

  12. [12]

    Flashtex: Fast relightable mesh texturing with lightcontrolnet

    Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. Flashtex: Fast relightable mesh texturing with lightcontrolnet. InEuropean conference on computer vi- sion, pages 90–107. Springer, 2024

    2024

  13. [13]

    Dreammat: High-quality pbr material generation with geometry-and light-aware diffu- sion models.ACM Transactions on Graphics (TOG), 43(4):1–18, 2024

    Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, et al. Dreammat: High-quality pbr material generation with geometry-and light-aware diffu- sion models.ACM Transactions on Graphics (TOG), 43(4):1–18, 2024

    2024

  14. [14]

    Objaverse-xl: A universe of 10m+ 3d objects.Advances in Neural Information Processing Systems, 36:35799–35813, 2023

    Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Chris- tian Laforte, Vikram V oleti, Samir Yitzhak Gadre, et al. Objaverse-xl: A universe of 10m+ 3d objects.Advances in Neural Information Processing Systems, 36:35799–35813, 2023

    2023

  15. [15]

    Ma- terial anything: Generating materials for any 3d object via diffusion

    Xin Huang, Tengfei Wang, Ziwei Liu, and Qing Wang. Ma- terial anything: Generating materials for any 3d object via diffusion. InProceedings of the Computer Vision and Pat- tern Recognition Conference, pages 26556–26565, 2025

    2025

  16. [16]

    Tryoffdiff: Virtual-try-off via high-fidelity gar- ment reconstruction using diffusion models.arXiv preprint arXiv:2411.18350, 2024

    Riza Velioglu, Petra Bevandic, Robin Chan, and Barbara Hammer. Tryoffdiff: Virtual-try-off via high-fidelity gar- ment reconstruction using diffusion models.arXiv preprint arXiv:2411.18350, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    One model for all: Partial diffusion for unified try-on and try-off in any pose.arXiv preprint arXiv:2508.04559, 2025

    Jinxi Liu, Zijian He, Guangrun Wang, Guanbin Li, and Liang Lin. One model for all: Partial diffusion for unified try-on and try-off in any pose.arXiv preprint arXiv:2508.04559, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    V oost: A unified and scalable diffusion transformer for bidirectional virtual try- on and try-off.arXiv preprint arXiv:2508.04825, 2025

    Seungyong Lee and Jeong-gi Kwak. V oost: A unified and scalable diffusion transformer for bidirectional virtual try- on and try-off.arXiv preprint arXiv:2508.04825, 2025

    arXiv 2025

  19. [19]

    Learning to transfer texture from clothing images to 3d humans

    Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7023–7034, 2020

    2020

  20. [20]

    Cloth2tex: A customized cloth texture generation pipeline for 3d vir- tual try-on

    Daiheng Gao, Xu Chen, Xindi Zhang, Qi Wang, Ke Sun, Bang Zhang, Liefeng Bo, and Qixing Huang. Cloth2tex: A customized cloth texture generation pipeline for 3d vir- tual try-on. In2024 international conference on 3D vision (3DV), pages 602–611. IEEE, 2024

    2024

  21. [21]

    Garment3dgen: 3d gar- ment stylization and texture generation

    Nikolaos Sarafianos, Tuur Stuyck, Xiaoyu Xiang, Yilei Li, Jovan Popovic, and Rakesh Ranjan. Garment3dgen: 3d gar- ment stylization and texture generation. In2025 Interna- tional Conference on 3D Vision (3DV), pages 1382–1393. IEEE, 2025

    2025

  22. [22]

    Materialmvp: Illumination- invariant material generation via multi-view pbr diffusion

    Zebin He, Mingxin Yang, Shuhui Yang, Yixuan Tang, Tao Wang, Kaihao Zhang, Guanying Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, et al. Materialmvp: Illumination- invariant material generation via multi-view pbr diffusion. arXiv preprint arXiv:2503.10289, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Active view selector: Fast and accurate active view selection with cross reference image quality assess- ment.arXiv preprint arXiv:2506.19844, 2025

    Zirui Wang, Yash Bhalgat, Ruining Li, and Victor Adrian Prisacariu. Active view selector: Fast and accurate active view selection with cross reference image quality assess- ment.arXiv preprint arXiv:2506.19844, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  25. [25]

    Deepfashion2: A versatile bench- mark for detection, pose estimation, segmentation and re-identification of clothing images

    Yuying Ge, Ruimao Zhang, Xiaogang Wang, Xiaoou Tang, and Ping Luo. Deepfashion2: A versatile bench- mark for detection, pose estimation, segmentation and re-identification of clothing images. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5337–5345, 2019. 9

    2019

  26. [26]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    4d-dress: A 4d dataset of real-world human clothing with semantic annotations

    Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 550–560, 2024. 10 A Implementation Details A.1 MR Rectification for Cross-p...

    2024