SMPL Normal Map Is All You Need for Single-view Textured Human Reconstruction

Gangjian Zhang; Hao Wang; Jianfeng Zhang; Nanjie Yao; Wenhao Shen; Xuanmeng Zhang; Yu Feng

REVIEW 3 major objections 5 minor 28 references

SMPL Normal Map Is All You Need for Single-view Textured Human Reconstruction

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper seeks to establish that four orthogonal SMPL normal maps, injected as guidance and constraint into a fine-tuned Gaussian reconstruction model, are sufficient for accurate single-view textured human reconstruction without diffusion

desk verdict A credible engineering contribution that adapts LGM to single-view human reconstruction with SMPL normal maps, reporting large benchmark gains, but it has a suspect LGM baseline and an untested training/inference SMPL gap. read the letter →

arxiv 2506.12793 v1 pith:TYPFLZ2R submitted 2025-06-15 cs.CV

Wenhao Shen , Gangjian Zhang , Jianfeng Zhang , Yu Feng , Nanjie Yao , Xuanmeng Zhang , Hao Wang This is my paper

classification cs.CV

keywords single-view3DhumanreconstructionSMPLnormalmapsGaussianSplattinglargemodelSEHRframeworktexturedavatarinvisiblebodypartmapguidance

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

The paper is trying to establish that a cheap parametric body prior—four orthogonal SMPL normal maps—is enough to make a large general-object 3D reconstruction model work for single-view textured human reconstruction. SEHR fine-tunes the large multi-view Gaussian model LGM with a single front-view RGB input, adds an auxiliary network that injects SMPL normal-map features into the main feature map, and trains the model to also predict SMPL normal Gaussians for the back and side views. If this is right, accurate clothed human avatars can be produced in roughly 35 seconds in one forward pass, avoiding the hallucinated patterns and Janus artifacts that appear when Stable Diffusion is used for novel views. Across the CustomHumans and CAPE benchmarks, the reported gains over the previous state of the art include roughly 30% lower Chamfer distance on CustomHumans.

What carries the argument

The load-bearing object is the SMPL normal map: a 2D rendering of surface normals from a parametric body mesh, taken from front, back, left, and right. The two mechanisms built on it are SNMG, which runs these maps through an auxiliary U-Net and residual-adds their features to the low-resolution Gaussian feature map of the main branch, and SNMC, which predicts a second set of Gaussians whose per-Gaussian features encode direction instead of color, renders the back, left, and right normal views from them, and supervises those renderings with the SMPL normal maps. The whole pipeline is supervised through differentiable 3D Gaussian Splatting rendering, so the SMPL shape signal propagates into the positions, scales, and rotations of the final Gaussians.

What would settle it

Run SEHR on a set of monocular images with deliberately degraded SMPL fits—for example, loose dresses, long hair, or hand poses that SMPL does not model—and compare reconstruction error against tight-clothing cases. If the Chamfer distance and f-score stay flat while the SMPL registration error grows, the claimed dependence on SMPL normal maps is falsified; if they degrade in step with the SMPL error, the prior is genuinely carrying the geometry.

Watch

Extended reading notes

Core claim

The central discovery is that SMPL normal maps carry enough human body structure to bridge the gap between generic object-level large reconstruction models and human-specific reconstruction, despite the scarcity of 3D human training data. SEHR does this in one forward pass: an off-the-shelf monocular SMPL estimator provides a mesh, four orthogonal normal maps are rendered from it, and those maps are used twice—once as residual feature guidance in the U-Net that produces the Gaussians, and once as a constraint that forces the model to render correct non-front normal views from an extra set of direction-feature Gaussians. The paper reports consistent improvements over feed-forward implicit-function methods and over diffusion-based SiTH, and attributes them to the shape information hidden in the normal maps rather than to larger datasets or more parameters.

Load-bearing premise

The method assumes that SMPL normal maps are an accurate proxy for the true clothed body geometry; at inference the SMPL mesh comes from a monocular estimator, so loose clothing, hair, or unusual poses can pull the guided Gaussians toward the wrong shape.

Editorial extensions

If this is right

Single-view human reconstruction no longer requires a multi-view diffusion stage: one RGB image plus four SMPL normal maps yields a textured avatar in a single feed-forward pass, in about 35 seconds per subject.
Large general-object reconstruction models can be redirected to humans with only a few thousand human scans, because the parametric prior supplies the category-specific structure the generic model lacks.
Invisible back and side geometry stops being a pure hallucination problem: SNMC gives the model a concrete normal-map target to satisfy, improving Chamfer distance and normal consistency on out-of-distribution subjects.
The output format, 3D Gaussians, can be converted to meshes, so the improved geometry carries over to existing mesh-based pipelines for evaluation and downstream use.

Reading between the lines

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

The paper does not test how sensitive the method is to SMPL estimation error; a natural extension is to corrupt or jitter the inference-time normal maps and measure the drop, which would reveal how much of the gain comes from the prior as opposed to the estimator.
The same recipe—a parametric normal-map prior injected into a large reconstruction model with a normal-map constraint—could transfer to other categories with strong parametric models, such as hands, faces, or quadrupeds, where training data are also scarce.
The feed-forward speed suggests interactive avatar creation is feasible, but the paper does not address whether the reconstructed Gaussians can be re-posed or animated; adding a canonicalization or skinning step would be a natural follow-up.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Worth a look if you work on single-view human reconstruction. The paper does something concrete: it takes the object-level LGM model, finetunes it on human data with a single-view input, and injects SMPL normal maps through an auxiliary U-Net (SNMG) plus a supervision head on non-front normal Gaussians (SNMC). That is a legitimate, non-obvious adaptation, and the reported numbers are strong: on CustomHumans, CD P2S/S2P drop to 1.262/1.666 versus SiTH's 1.871/2.045, NC goes from 0.826 to 0.848, f-Score from 37.0 to 48.0, and inference is 35 seconds. The ablations are internally consistent: adding more normal-map views helps, and the three-view SNMC beats the four-view variant in a way that makes sense. The writing is clear and the method section is reproducible in principle.

The main soft spots are verification-related, not conceptual. First, the LGM row in Table I is numerically identical to the SHGM baseline in Table III (2.085/2.851 on CustomHumans). That means the authors likely ran their own single-view adaptation and labeled it LGM, rather than the actual multi-view LGM pipeline. That is a baseline reporting error, and it inflates the apparent gain over the original LGM. Second, there are no error bars or significance tests, so we do not know whether the margins over SiTH are stable across random seeds. Third, the training/inference gap is real: training uses ground-truth SMPL normal maps from THuman2.0, while inference uses SMPLer-X estimates. The stress-test concern is legitimate—loose clothing, hair, or pose error could degrade the guidance—but I would not call it load-bearing, because the ablations show the method works with imperfect SMPL too (SMPLer-X is not perfect on CAPE either, and the gains persist). Still, the paper should have included a sensitivity analysis or at least a qualitative note on this.

The title is overclaiming: SMPL normal maps are not "all you need"; the method also uses the RGB image and the pretrained LGM weights. That is a marketing problem, not a scientific one.

Overall: a solid empirical paper with a real but incremental contribution, one clear reporting flaw, and one untested robustness assumption. It deserves peer review, not desk rejection. A serious referee should ask for the code, the corrected LGM baseline, error bars, and a brief SMPL-robustness experiment.

I would bring it to a reading group, and I would cite it if I needed a reference for single-view human reconstruction with Gaussian splatting.

Referee Report

3 major / 5 minor

Summary. The paper proposes SEHR, a single-view textured human reconstruction framework built by fine-tuning the large multi-view Gaussian model LGM with SMPL normal maps as a human geometry prior. Two components are introduced: SNMG, which injects four orthogonal SMPL normal maps into an auxiliary U-Net and fuses the resulting features into the main Gaussian reconstruction branch via residual connections, and SNMC, which predicts an extra set of SMPL normal Gaussians for non-front views and supervises their rendered normal maps against ground-truth SMPL normal maps. The model is trained on THuman2.0 and evaluated on CAPE and CustomHumans, where it reports improved Chamfer Distance, Normal Consistency, f-Score, and LPIPS over existing methods, with a 35-second inference time.

Significance. If the reported results hold, SEHR is a practically valuable contribution: it demonstrates that a large pretrained object-level Gaussian reconstruction model can be adapted to human reconstruction with a lightweight parametric-body prior, avoiding the multi-view diffusion front-end that causes hallucination and Janus problems in many prior systems. The formulation is clean, the ablation study is internally consistent, and the reported inference speed is a clear practical advantage. The paper also makes a useful conceptual point that SMPL normal maps, often used only as intermediate supervision, can serve as a direct conditioning signal for feed-forward Gaussian prediction. However, the strength of these contributions is currently undermined by a baseline-comparison issue and by an unexamined train/inference gap in the source of SMPL normal maps, so the quantitative claims should be treated as provisional until these points are addressed.

major comments (3)

[Table I and Table III] The LGM row in Table I on CustomHumans reports CD 2.085/2.851, NC 0.794, and f-Score 30.389, which are exactly the same values reported for SHGM in Table III. This strongly suggests that the 'LGM' baseline is actually the authors' own single-view SHGM baseline rather than the original LGM model, which uses a multi-view diffusion model to generate four orthogonal images. If that is the case, the comparison against LGM is circular and the claim that SEHR improves over LGM is not established. Please clarify how the LGM baseline was obtained, and if the numbers are indeed SHGM, rerun the original LGM under the same evaluation protocol or remove the LGM comparison.
[Section IV.A and Section III.C] The training and inference pipelines use different sources of SMPL normal maps: training renders them from ground-truth SMPL meshes in THuman2.0, while inference renders them from SMPL meshes estimated by SMPLer-X. The paper provides no analysis of how SMPL estimation error, loose clothing, or hair affects SNMG and SNMC, even though both components feed these normal maps directly into the Gaussian prediction and supervision. This is load-bearing because the claimed advantage over diffusion-based baselines rests on the reliability of the SMPL prior. Please add a sensitivity study, for example comparing inference with ground-truth SMPL meshes versus SMPLer-X estimates on a held-out subset, perturbing SMPL pose and shape parameters, or stratifying results by clothing tightness.
[Table I and Table IV] The paper reports no error bars, multiple seeds, or statistical tests. Some of the headline margins are small, notably on CAPE where the CD P2S improvement over SiTH is 0.135 cm and the f-Score improvement is 1.289, and in Table IV where several SNMC ablations are within 0.05 cm of each other. Without variance estimates it is impossible to tell whether these differences are meaningful. Please report mean and standard deviation over at least three training runs, or otherwise justify why a single run is sufficient for the stated claims.

minor comments (5)

[Equation (2)] The normal-map supervision uses LPIPS, which is a perceptual metric trained on natural RGB images. Applying it to SMPL normal maps, whose channel values are surface normals rather than colors, may not be meaningful; please justify this choice or report results with the LPIPS term removed.
[Section III.D] The description of SMPL normal Gaussians is brief: it states that the color feature is replaced by a direction feature, but it is not clear how these directions are defined, how they are rendered into normal maps, or how the rendered normal maps are aligned with the ground-truth SMPL normal maps. Please expand this part.
[Table IV caption] The caption uses both 'HNMC' and the text uses 'HNNC' for the human normal map constraint; please use consistent terminology and define each acronym at first use.
[Section IV.A] The dataset description is too terse: please specify the number of training and test subjects, the exact CAPE and CustomHumans splits used, and how the ground-truth SMPL normal maps are rendered (which SMPL implementation, which renderer).
[General] There are several typographical issues, including 'novelSMPL' in the introduction and inconsistent spacing in equations; a careful proofread is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SEHR is a supervised feed-forward method evaluated on held-out benchmarks; the SMPL prior is an external assumption rather than a fitted input.

full rationale

SEHR's derivation chain is a supervised learning pipeline, not a closed-form derivation. The training losses in Eqs. (1) and (2) supervise rendered RGB/mask outputs and rendered SMPL normal maps against ground-truth images and ground-truth SMPL normal maps obtained from the GT SMPL mesh (Section IV.A). SMPL normal maps enter as input features (SNMG) and as an auxiliary supervision target (SNMC), while the final evaluation is Chamfer Distance, normal consistency, and f-score against held-out CAPE and CustomHumans scans. No parameter is fitted to the test sets, and no reported improvement is an algebraic restatement of an input. The inference-time use of SMPLer-X estimates versus ground-truth SMPL during training is a real domain-gap robustness concern, but it is not circular reasoning: the SMPL body prior is an external assumption about human shape, not a quantity defined by the target clothed reconstruction. The cited prior work (LGM, SMPL, SMPLer-X, Splatter Image) is external; no load-bearing argument reduces to a self-citation or to a uniqueness theorem imported from the authors. The central claim therefore has independent empirical content, and no circular step can be identified from the paper's own equations or citations.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

This is a supervised deep learning method, not a derivation, so no scientific constants are introduced. The ledger lists the main hand-chosen design choices and the external assumptions (SMPL prior, SMPLer-X, LGM transfer) that the central claim depends on. The sole invented entity is the SMPL normal Gaussian representation used for supervision; it has no independent validation outside the paper.

free parameters (5)

Network weights of SHGM, SNMG, and SNMC
All model parameters are learned by fine-tuning on THuman2.0; the paper does not report parameter counts or training curves.
Learning rate and optimizer = AdamW with lr=1e-4
Hand-chosen training hyperparameters reported in Section IV.A without ablations.
Loss weighting = w_rgb = w_norm = 1
The total loss is Lrgb + Lnorm with no weighting scheme (Eq. 3), an unexamined choice.
Number of SMPL normal map views = 4 (front, back, left, right)
Table III shows performance improves with more views; four is the selected design without a formal optimum.
Rendered supervision views = 8 (1 input + 7 novel)
Chosen in Section IV.A; no ablation is shown for this number.

assumptions (5)

domain assumption SMPL is a valid parametric body prior for clothed human reconstruction.
SMPL models the naked body, yet its normal maps are used to guide clothed surface reconstruction in Section III.C.
domain assumption SMPLer-X provides sufficiently accurate SMPL meshes from monocular images for inference.
Section III.C relies on SMPLer-X at inference; the paper does not analyze the effect of pose or shape estimation errors.
domain assumption Ground-truth SMPL normal maps are an appropriate supervision signal for learning invisible parts.
Section IV.A renders normal maps from ground-truth SMPL meshes; the mismatch with the true clothed surface is not addressed.
domain assumption LGM's pretrained object representation transfers to human bodies after fine-tuning.
Section I assumes LGM is rich in 3D knowledge; the paper does not quantify how much of the improvement comes from this transfer.
standard math Differentiable Gaussian rasterization gives correct gradients for RGB, mask, and normal losses.
The method uses the standard 3D Gaussian splatting renderer from LGM (Section III.A).

invented entities (1)

SMPL normal Gaussians (Theta')
purpose: A set of 3D Gaussians storing normal directions instead of colors, rendered to supervise non-front views in SNMC.
This is an internal representation introduced by SNMC; its value is evidenced only by the benchmark improvements reported in this paper, with no external falsifiable prediction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SMPL Normal Map Is All You Need for Single-view Textured Human Reconstruction." pith.science (2026). https://pith.science/paper/TYPFLZ2R

@misc{pith2026250612793,
  author       = {Pith},
  title        = {Pith review of: SMPL Normal Map Is All You Need for Single-view Textured Human Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TYPFLZ2R}},
  note         = {Machine review of arXiv:2506.12793}
}

read the original abstract

Single-view textured human reconstruction aims to reconstruct a clothed 3D digital human by inputting a monocular 2D image. Existing approaches include feed-forward methods, limited by scarce 3D human data, and diffusion-based methods, prone to erroneous 2D hallucinations. To address these issues, we propose a novel SMPL normal map Equipped 3D Human Reconstruction (SEHR) framework, integrating a pretrained large 3D reconstruction model with human geometry prior. SEHR performs single-view human reconstruction without using a preset diffusion model in one forward propagation. Concretely, SEHR consists of two key components: SMPL Normal Map Guidance (SNMG) and SMPL Normal Map Constraint (SNMC). SNMG incorporates SMPL normal maps into an auxiliary network to provide improved body shape guidance. SNMC enhances invisible body parts by constraining the model to predict an extra SMPL normal Gaussians. Extensive experiments on two benchmark datasets demonstrate that SEHR outperforms existing state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2506.12793 by the authors.

**Figure 1.** Overview of SEHR. Based on the SHGM model, SEHR introduces an extra branch to extract the human prior in the orthogonal SMPL normal maps and provide shape guidance for Gaussian learning, called SMPL Normal Map Guidance (SNMG). Additionally, we propose using the back, left, and right views of SMPL normal maps to constrain the hallucination of invisible body parts, named SMPL Normal Map Constraint (SNMC). shares the s… view at source ↗

**Figure 2.** Visualization comparison with SOTA methods. It is shown that our method has a better reconstruction effect on the details of the 3D human body, such as the fingers and edges of the clothes. Meanwhile, our method also has fewer hallucination artifacts compared to SiTH [7], which wrongly generates patterns that do not exist on clothes. Please zoom in for a better view. TABLE IV ABLATION OF SNMC. “HNMC(HUMAN NORMAL MAP… view at source ↗

**Figure 3.** illustrates the quality enhancement brought by SNMG. The input image remains unchanged across comparisons, emphasizing the differences in human shape between [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Visual ablation of SNMC. We compare the methods with and without SNMC. While SNMG significantly enhances the overall shape of the reconstructed 3D mesh, certain subtle details, such as the shoulder hull in the first case, remain slightly inaccurate. In contrast, SNMC f…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 22 canonical work pages

[1]

ICON: Implicit Clothed humans Obtained from Normals,

Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black, “ICON: Implicit Clothed humans Obtained from Normals,” inCVPR, June 2022

work page 2022
[2]

ECON: Explicit Clothed humans Optimized via Normal integration,

Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black, “ECON: Explicit Clothed humans Optimized via Normal integration,” inCVPR, June 2023

work page 2023
[3]

Global-correlated 3d-decoupling transformer for clothed avatar recon- struction,

Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang, “Global-correlated 3d-decoupling transformer for clothed avatar recon- struction,”NeurIPS, 2024

work page 2024
[4]

Sifu: Side-view condi- tioned implicit function for real-world usable clothed human reconstruc- tion,

Zechuan Zhang, Zongxin Yang, and Yi Yang, “Sifu: Side-view condi- tioned implicit function for real-world usable clothed human reconstruc- tion,” inCVPR, June 2024

work page 2024
[5]

Smpl: a skinned multi-person linear model,

Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black, “Smpl: a skinned multi-person linear model,” TOG, 2015

work page 2015
[6]

Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors,

Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu, “Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors,” inCVPR, June 2021

work page 2021
[7]

Sith: Single-view textured human reconstruction with image-conditioned diffusion,

Hsuan-I Ho, Jie Song, and Otmar Hilliges, “Sith: Single-view textured human reconstruction with image-conditioned diffusion,” inCVPR, 2024

work page 2024
[8]

TeCH: Text-guided Reconstruction of Lifelike Clothed Humans,

Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Jiaxiang Tang, Deng Cai, and Justus Thies, “TeCH: Text-guided Reconstruction of Lifelike Clothed Humans,” in3DV, 2024

work page 2024

Show all 28 references

[9]

Adding conditional control to text-to-image diffusion models,

Lvmin Zhang, Anyi Rao, and Maneesh Agrawala, “Adding conditional control to text-to-image diffusion models,” inICCV, 2023

2023
[10]

Dream- fusion: Text-to-3d using 2d diffusion,

Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall, “Dream- fusion: Text-to-3d using 2d diffusion,”arXiv:2209.14988, 2022

2022 arXiv
[11]

High-resolution image synthesis with latent diffu- sion models,

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer, “High-resolution image synthesis with latent diffu- sion models,” inCVPR, June 2022

2022
[12]

Lgm: Large multi-view gaussian model for high- resolution 3d content creation,

Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu, “Lgm: Large multi-view gaussian model for high- resolution 3d content creation,”arXiv:2402.05054, 2024

2024 arXiv
[13]

Lrm: Large reconstruction model for single image to 3d,

Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan, “Lrm: Large reconstruction model for single image to 3d,”arXiv:2311.04400, 2023

2023 arXiv
[14]

U-net: Con- volutional networks for biomedical image segmentation,

Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Con- volutional networks for biomedical image segmentation,” inMICCAI. Springer, 2015

2015
[15]

3d gaussian splatting for real-time radiance field rendering,

Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis, “3d gaussian splatting for real-time radiance field rendering,” TOG, July 2023

2023
[16]

Expres- sive body capture: 3d hands, face, and body from a single image,

Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black, “Expres- sive body capture: 3d hands, face, and body from a single image,” in CVPR, 2019

2019
[17]

Smpler-x: Scaling up expressive human pose and shape estimation,

Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al., “Smpler-x: Scaling up expressive human pose and shape estimation,” NeurIPS, 2024

2024
[18]

Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization,

Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li, “Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization,” inICCV, October 2019

2019
[19]

Pi- fuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization,

Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo, “Pi- fuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization,” inCVPR, 2020

2020
[20]

Arch: Animatable reconstruction of clothed humans,

Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung, “Arch: Animatable reconstruction of clothed humans,” inCVPR, 2020

2020
[21]

High-fidelity 3d human digitization from single 2k resolution images,

Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon, “High-fidelity 3d human digitization from single 2k resolution images,” inCVPR, 2023

2023
[22]

Photore- alistic monocular 3d reconstruction of humans wearing clothing,

Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu, “Photore- alistic monocular 3d reconstruction of humans wearing clothing,” in CVPR, 2022

2022
[23]

Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model,

Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al., “Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model,” arXiv:2311.09217, 2023

2023 arXiv
[24]

Splatter image: Ultra-fast single-view 3d reconstruction,

Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi, “Splatter image: Ultra-fast single-view 3d reconstruction,” arXiv:2312.13150, 2023

2023 arXiv
[25]

Learning locally editable virtual humans,

Hsuan-I Ho, Lixin Xue, Jie Song, and Otmar Hilliges, “Learning locally editable virtual humans,” inCVPR, 2023

2023
[26]

Pamir: Paramet- ric model-conditioned implicit representation for image-based human reconstruction,

Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai, “Pamir: Paramet- ric model-conditioned implicit representation for image-based human reconstruction,”TPAMI, 2021

2021
[27]

Fof: Learning fourier occupancy field for monocular real-time human recon- struction,

Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li, “Fof: Learning fourier occupancy field for monocular real-time human recon- struction,” inNeurIPS, 2022

2022
[28]

Learning to dress 3d people in generative clothing,

Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons- Moll, Siyu Tang, and Michael J Black, “Learning to dress 3d people in generative clothing,” inCVPR, 2020

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

ICON: Implicit Clothed humans Obtained from Normals,

Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black, “ICON: Implicit Clothed humans Obtained from Normals,” inCVPR, June 2022

work page 2022

[2] [2]

ECON: Explicit Clothed humans Optimized via Normal integration,

Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black, “ECON: Explicit Clothed humans Optimized via Normal integration,” inCVPR, June 2023

work page 2023

[3] [3]

Global-correlated 3d-decoupling transformer for clothed avatar recon- struction,

Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang, “Global-correlated 3d-decoupling transformer for clothed avatar recon- struction,”NeurIPS, 2024

work page 2024

[4] [4]

Sifu: Side-view condi- tioned implicit function for real-world usable clothed human reconstruc- tion,

Zechuan Zhang, Zongxin Yang, and Yi Yang, “Sifu: Side-view condi- tioned implicit function for real-world usable clothed human reconstruc- tion,” inCVPR, June 2024

work page 2024

[5] [5]

Smpl: a skinned multi-person linear model,

Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black, “Smpl: a skinned multi-person linear model,” TOG, 2015

work page 2015

[6] [6]

Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors,

Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu, “Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors,” inCVPR, June 2021

work page 2021

[7] [7]

Sith: Single-view textured human reconstruction with image-conditioned diffusion,

Hsuan-I Ho, Jie Song, and Otmar Hilliges, “Sith: Single-view textured human reconstruction with image-conditioned diffusion,” inCVPR, 2024

work page 2024

[8] [8]

TeCH: Text-guided Reconstruction of Lifelike Clothed Humans,

Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Jiaxiang Tang, Deng Cai, and Justus Thies, “TeCH: Text-guided Reconstruction of Lifelike Clothed Humans,” in3DV, 2024

work page 2024