MultiGO++: Monocular 3D Clothed Human Reconstruction via Geometry-Texture Collaboration

Nanjie Yao; Gangjian Zhang; Wenhao Shen; Jian Shu; Yu Feng; Hao Wang

cs.CVMar 5, 2026

MultiGO++: Monocular 3D Clothed Human Reconstruction via Geometry-Texture Collaboration

Authors: Nanjie Yao, Gangjian Zhang, Wenhao Shen, Jian Shu, Yu Feng, Hao Wang

Organizations: The Hong Kong University of Science and Technology (Guangzhou), Guangzhou 511442, China · Nanyang Technological University, Singapore 639798

Abstract

Monocular 3D clothed human reconstruction aims to generate a complete and realistic textured 3D avatar from a single image. Existing methods are commonly trained under multi-view supervision with annotated geometric priors, and during inference, these priors are estimated by the pre-trained network from the monocular input. These methods are constrained by three key limitations: texturally by unavailability of training data, geometrically by inaccurate external priors, and systematically by biased single-modality supervision, all leading to suboptimal reconstruction. To address these issues, we propose a novel reconstruction framework, named MultiGO++, which achieves effective systematic geometry-texture collaboration. It consists of three core parts: (1) A multi-source texture synthesis strategy that constructs 15,000+ 3D textured human scans to improve the performance on texture quality estimation in challenge scenarios; (2) A region-aware shape extraction module that extracts and interacts features of each body region to obtain geometry information and a Fourier geometry encoder that mitigates the modality gap to achieve effective geometry learning; (3) A dual reconstruction U-Net that leverages geometry-texture collaborative features to refine and generate high-fidelity textured 3D human meshes. Extensive experiments on two benchmarks and many in-the-wild cases show the superiority of our method over state-of-the-art approaches.

Figures & tables

Fig. 2: Method Overview. Our framework integrates three core components: Texturally, we employ a multi-source texture synthesis strategy to generate diverse synthetic data for training, along with a lightweight texture encoder for effective feature extraction. Geometrically, we introduce a Region-aware Shape Extraction Module that enhances human shape extraction through part-based feature interaction, utilizing Self-Attention (SA), Cross-Attention (CA), and Feed-Forward Networks (FFN). This is coupled with a Fourier Geometry Encoder to bridge the modality gap for efficient geometric learning. Systematically, we propose a Dual Reconstruction U-Net that utilizes feature residuals to balance geometric and texture features, enabling mutual enhancement across modalities. Additionally, to refine 3D mesh quality and extraction efficiency, we design a Gaussian-enhanced remeshing strategy supervised by the generated normal Gaussian avatar.

Fig. 3: Multi-source Texture Synthesis Strategy. The proposed multi-source texture synthesis strategy leverages X-to-3D models and multimodal LLM data screening to generate high-quality training data for enhanced texture estimation.

Fig. 4: Detailed Architecture of Fourier Geometry Encoder. To achieve effective geometry learning, we achieve better fusion of the heterogeneous modalities of the 3D geometry prior and 2D images. We propose interpolating the Fourier features of 3D occluded points and mapping them from three different angles into the same 2D space as the image features.

Fig. 5: Qualitative comparisons on in-the-wild images featuring loose clothing. While other SOTA methods struggle to accurately reconstruct the challenging geometries of loose garments, our approach faithfully reproduces high-fidelity wrinkles and intricate textures. Please zoom in for a detailed view.

Methods	Publication	CustomHuman [ 13 ]			THuman3.0 [ 47 ]
Methods	Publication	CD: P-to-S / S-to-P $(\mathrm{cm})\downarrow$	NC $\uparrow$	F-score $\uparrow$	CD: P-to-S / S-to-P $(\mathrm{cm})\downarrow$	NC $\uparrow$	F-score $\uparrow$
PIFu [ 43 ]	ICCV 2019	$2.965/3.108$	0.765	25.708	$2.176/2.452$	0.773	34.194
ICON [ 56 ]	CVPR 2022	$2.441/2.823$	0.785	29.144	$2.368/2.776$	0.754	27.434
ECON [ 55 ]	CVPR 2023	$2.196/2.340$	0.801	33.292	$2.201/2.271$	0.783	33.223
GTA [ 72 ]	NeurIPS 2023	$2.404/2.726$	0.790	29.907	$2.416/2.652$	0.768	29.257
VS [ 32 ]	CVPR 2024	$2.518/2.993$	0.780	26.791	$2.526/2.942$	0.753	26.344

TABLE I: Comparison of Human Geometry with SOTA methods. The best and second results are highlighted with bold and underline respectively. Arrow

\uparrow

\downarrow

means higher/lower is better. Grey background represents the test set is used as training data, and these methods are excluded from the ranking. “ † ” indicates the models trained on more commercial or synthesis data.

Methods	CustomHuman			THuman3.0
Methods	LPIPS: F/B $\downarrow$	SSIM: F/B $\uparrow$	PSNR: F/B $\uparrow$	LPIPS: F/B $\downarrow$	SSIM: F/B $\uparrow$	PSNR: F/B $\uparrow$
PIFu [ 43 ]	$0.0792/0.0966$	$0.8965/0.8742$	$18.141/16.721$	$0.0706/0.0849$	$0.9242/0.9007$	$20.104/17.926$
ICON [ 56 ]	$0.0714/-$	$0.8975/-$	$18.614/-$	$0.0602/-$	$0.9287/-$	$21.126/-$
ECON [ 55 ]	$0.0777/-$	$0.8870/-$	$18.437/-$	$0.0638/-$	$0.9258/-$	$20.951/-$
GTA [ 73 ]	$0.0730/0.0891$	$0.9003/0.8923$	$18.790/18.229$	$0.0633/0.0770$	$0.9298/0.9275$	$21.113/20.497$
SIFU [ 74 ]	$0.0682/0.0880$	$0.9018/0.8907$	$18.710/18.114$	$0.0594/0.0764$	$0.9307/0.9245$	$21.103/20.351$

TABLE II: Comparison of Human Texture with SOTA methods. For comparison, we render the textured 3D human reconstruction results of these methods in front view and back view, represented by “F/B” symbols. Note that only partial methods reconstruct 3D human texture, and ICON and ECON only reconstruct the front-view texture.

Fig. 6: Visual Comparisons with SOTA methods. Previous SOTA approaches always struggle with recovering the correct human pose, shape, and fine-grained geometry and texture details. For a detailed view, please zoom in on the image.

Method	ICON	ECON	GTA	VS	HiLo
Infer. Time	20s	8.5min	4.5min	20min	1.5mins
Method	SIFU	SiTH	R 2 Human	HumanRef	PSHuman
Infer. Time	6mins	2mins	2s	2h	50s
Method	H3Diff.	MultiGO	MultiGO++
Infer. Time	2mins	0.6s	0.7s
M.E. Time	12min	3min	1min

TABLE III: Evaluation on Computational Efficiency. Our MultiGO++ achieves the fastest mesh extraction while maintaining highly competitive inference time.

Fig. 7: Qualitative comparisons on in-the-wild images featuring challenging poses. Compared to existing approaches, our method robustly reconstructs geometrically accurate limb structures and preserves fine-grained facial expressions, while introducing significantly fewer visual artifacts. Please zoom in for a detailed view.

Fig. 8: Quality Evaluation on Synthetic data. We show randomly selected data from HuGe-100K and our dataset, where locations with cross-view inconsistencies are marked with red boxes. Please zoom in for a detailed view.

Methods	Section	CustomHuman [ 13 ]			THuman3.0 [ 47 ]
Methods	Section	CD: P-to-S / S-to-P $(\mathrm{cm})\downarrow$	NC $\uparrow$	F-score $\uparrow$	CD: P-to-S / S-to-P $(\mathrm{cm})\downarrow$	NC $\uparrow$	F-score $\uparrow$
w/ $Data_{Com.+Syn.}$	Texture	$1.402/1.562$	0.865	47.208	$1.173/1.299$	0.850	53.480
w/ $Data_{Com.}$		$1.476/1.632$	0.860	45.972	$1.214/1.318$	0.840	52.545
w/o $Data.$		$1.481/1.652$	0.859	45.038	$1.237/1.406$	0.842	51.012
w/ 3-view Proj.	Geometry	$1.462/1.658$	0.858	45.081	$1.221/1.348$	0.841	52.128
w/ 2-view Proj.		$1.546/1.761$	0.845	42.884	$1.267/1.831$	0.834	50.621

TABLE IV: Ablation Study on Geometry Accuracy. Texture : To evaluate the effectiveness of texture synthesis strategy proposed in section on model performance by comparing results using synthetic data against high-quality commercial training data. Geometry : we conduct an ablation study by replacing the Region-aware Shape Extraction Module (RSEM) during inference and evaluate the effectiveness of the Fourier Geometry Encoder (FGE) for reconstructing human geometry by comparing model performance with and without the 2D projection of 3D features. System : We assess the contributions of the dual U-Net by separately ablated normal U-Net and the Gaussian enhanced remeshing strategy.

CustomHuman	LPIPS: F/B $\downarrow$	SSIM: F/B $\uparrow$	PSNR: F/B $\uparrow$
w/ $Data_{Com.+Syn.}$	$0.0372/0.0599$	$0.9679/0.9475$	$23.646/21.313$
w/ $Data_{Com.}$	$0.0407/0.0634$	$0.9666/0.9451$	$23.412/21.012$
w/o $Data$	$0.0414/0.0641$	$0.9654/0.9423$	$23.122/20.865$
w/ Simplify	$0.0443/0.0667$	$0.9588/0.9364$	$22.569/20.456$
w/ HMR2.0	$0.0412/0.0610$	$0.9621/0.9422$	$22.912/20.707$
w/o FGE	$0.0462/0.0689$	$0.9502/0.9312$	$22.437/20.347$

TABLE V: Ablation Study on Texture Quality. We demonstrate the reconstructed texture quality is also improved with our proposed region-aware shape extraction module, Fourier geometry encoder and multi-source texture synthesis strategy.

Fig. 9: Visual Ablation. ( left ) Synthetic data improves the model performance on human texture estimation. ( middle ) The Fourier geometry encoder enhances the better feature fusion between 2D and 3D, enabling the overall human poses close to the ground truth. ( right ) The proposed region-aware shape extraction module improves the pose correctness. The remeshing process improves the geometry details in extracted meshes. Please zoom in to observe details.

3D Reconstruction

3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.

See all 955 papers

OAHuman: Occlusion-Aware 3D Human Reconstruction from Monocular Images
Mar 15, 2026Yuanwang Yang, Hongliang Liu, Muxin Zhang +43D Reconstruction Monocular 3D Reconstruction
College of Intelligence and Computing, Tianjin University, Tianjin 300350, China · College of Artificial Intelligence, Beijing University of Technology, Beijing 100124, China · School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China +1
Human Interaction-Aware 3D Reconstruction from a Single Image
Apr 7, 2026Gwanghyun Kim, Junghun James Kim, Suh Yoon Jeon +23D Human Reconstruction Single-View 3D Reconstruction
Dept. of Electrical and Computer Engineering, Seoul National University, Republic of Korea · IPAI, Seoul National University, Republic of Korea · INMC, Seoul National University, Republic of Korea
MuNet: A Mutualistic Network for Joint 3D Human Mesh Recovery and 3D Clothed Human Reconstruction from Single Images
May 25, 2026Yunqi Gao, Leyuan Liu, Yuhan Li +23D Reconstruction Human Mesh Recovery
National Engineering Research Center for E-Learning Central China Normal University, Wuhan, 430079, Hubei, China · National Engineering Research Center of Educational Big Data Central China Normal University, Wuhan, 430079, Hubei, China · School of Electronic Information and Communications Huazhong University of Science and Technology, Wuhan, 430073, Hubei, China +1

Mar 15, 2026J/K move · Enter open · S save

OAHuman: Occlusion-Aware 3D Human Reconstruction from Monocular Images

Yuanwang Yang, Hongliang Liu, Muxin Zhang, Nan Ma+3

College of Intelligence and Computing, Tianjin University, Tianjin 300350, China · College of Artificial Intelligence, Beijing University of Technology, Beijing 100124, China · School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China+1

3D Reconstruction Monocular 3D Reconstruction 3D Human Reconstruction Texture Synthesis

Open paper AskPDF arXiv

Monocular 3D human reconstruction in real-world scenarios remains highly challenging due to frequent occlusions from surrounding objects, people, or image truncation. Such occlusions lead to missing geometry and unreliable appearance cues, severely degrading the completeness and realism of reconstructed human models. Although recent neural implicit methods achieve impressive results on clean inputs, they struggle under occlusion due to entangled modeling of shape and texture. In this paper, we propose OAHuman, an occlusion-aware framework that explicitly decouples geometry reconstruction and texture synthesis for robust 3D human modeling from a single RGB image. The core innovation lies in the decoupling-perception paradigm, which addresses the fundamental issue of geometry-texture cross-contamination in occluded regions. Our framework ensures that geometry reconstruction is perceptually reinforced even in occluded areas, isolating it from texture interference. In parallel, texture synthesis is learned exclusively from visible regions, preventing texture errors from being transferred to the occluded areas. This decoupling approach enables OAHuman to achieve robust and high-fidelity reconstruction under occlusion, which has been a long-standing challenge in the field. Extensive experiments on occlusion-rich benchmarks demonstrate that OAHuman achieves superior performance in terms of structural completeness, surface detail, and texture realism, significantly improving monocular 3D human reconstruction under occlusion conditions.

Abstract

3D Reconstruction

Explore similar work

OAHuman: Occlusion-Aware 3D Human Reconstruction from Monocular Images

Human Interaction-Aware 3D Reconstruction from a Single Image

MuNet: A Mutualistic Network for Joint 3D Human Mesh Recovery and 3D Clothed Human Reconstruction from Single Images