Structure-Aware Fine-Grained Gaussian Splatting for Expressive Avatar Reconstruction

Yuze Su; Hongsong Wang; Jie Gui; Liang Wang

cs.CVApr 10, 2026

Structure-Aware Fine-Grained Gaussian Splatting for Expressive Avatar Reconstruction

Authors: Yuze Su, Hongsong Wang, Jie Gui, Liang Wang

Organizations: School of Cyber Science and Engineering, Southeast University, Nanjing 210096, China · School of Computer Science and Engineering, Southeast University, Nanjing 210096, China · Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China · Purple Mountain Laboratories, Nanjing 210000, China · Engineering Research Center of Blockchain Application, Supervision And Management (Southeast University), Ministry of Education, China · State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Beijing, China · Institute of Automation, Chinese Academy of Sciences (CASIA), Beijing, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China

Abstract

Reconstructing photorealistic and topology-aware human avatars from monocular videos remains a significant challenge in the fields of computer vision and graphics. While existing 3D human avatar modeling approaches can effectively capture body motion, they often fail to accurately model fine details such as hand movements and facial expressions. To address this, we propose Structure-aware Fine-grained Gaussian Splatting (SFGS), a novel method for reconstructing expressive and coherent full-body 3D human avatars from a monocular video sequence. The SFGS use both spatial-only triplane and time-aware hexplane to capture dynamic features across consecutive frames. A structure-aware gaussian module is designed to capture pose-dependent details in a spatially coherent manner and improve pose and texture expression. To better model hand deformations, we also propose a residual refinement module based on fine-grained hand reconstruction. Our method requires only a single-stage training and outperforms state-of-the-art baselines in both quantitative and qualitative evaluations, generating high-fidelity avatars with natural motion and fine details. The code is on Github: https://github.com/Su245811YZ/SFGS

Figures & tables

Figure 1: Limitation of conventional human gaussian splatting methods. The second row shows the error between the reconstructed 3D avatar from the corresponding input image in the first row and the 3D ground truth, with darker colors indicating larger reconstruction errors. The hand exhibits the largest reconstruction error across the entire body.

Figure 2: Our framework for creating animatable avatars from monocular videos. We first initialize a set of 3D Gaussians in the canonical space via sampling points from a SMPL-X mesh.The human Gaussians are parameterized by their mean locations and the time

t

in the canonical space, and adaptive feature fusion is performed using a HexPlane and a triplane to obtain the final representation

f

.Use MLPs to estimate the initial position offset, scale and color.To address the limitations of model-based reconstruction in capturing fine non-rigid deformations, we further introduce a structure-aware offset prediction module that explicitly incorporates joint priors.In addition, we adopt the more detailed MANO model to specifically reconstruct the hands, providing higher fidelity compared to the coarse hand modeling in SMPL-X.Finally, the avatar is animated based on driving signals, SMPL-X parameters, and camera configurations, and is rendered into screen space using 3D Gaussian Splatting (3DGS).

Figure 3: Motivated by the intuition that each Gaussian point should deform under the influence of its most relevant joint, we introduce a joint-aware feature $\mathbf{e}_{\text{joint}}$ to control its scale and positional deformation. A similar mechanism is employed to model the color offset.

Method	PSNR ↑	SSIM ↑	LPIPS ↓
NeuMan [ 13 ]	29.32	0.972	0.014
Vid2Avatar [ 8 ]	30.70	0.980	0.014
GaussianAvatar [ 10 ]	29.94	0.980	0.012
3DGS-Avatar [ 15 ]	28.99	0.974	0.016
Exavatar [ 24 ]	34.80	0.984	0.009
SFGS (ours)	35.34	0.985	0.009

Table 1: Quantitative comparison on the Neuman dataset. To align with prior work, we report the mean results over the sequences bike, citron, jogging, and seattle.

Methods	Seattle			Citron			Parking			Bike			Jogging			Lab
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Vid2Avatar [ 8 ]	16.90	0.51	0.27	15.96	0.59	0.28	18.51	0.65	0.26	12.44	0.39	0.54	16.36	0.46	0.30	15.99	0.62	0.34
NeuMan [ 13 ]	18.42	0.58	0.20	18.39	0.64	0.19	17.66	0.66	0.24	19.05	0.66	0.21	17.57	0.54	0.29	18.76	0.73	0.23
HUGS [ 16 ]	19.06	0.67	0.15	19.16	0.71	0.16	19.44	0.73	0.17	19.48	0.67	0.18	17.45	0.59	0.27	18.79	0.76	0.18
ExAvatar [ 24 ]	36.98	0.991	0.006	35.43	0.987	0.008	37.16	0.988	0.012	32.81	0.981	0.011	32.72	0.977	0.013	34.69	0.981	0.013
SFGS (ours)	37.58	0.991	0.006	35.49	0.986	0.009	37.50	0.988	0.010	34.75	0.983	0.012	33.55	0.978	0.013	34.85	0.987	0.011

Table 2: Quantitative comparison on six sequences of the Neuman dataset . The results of Exavatar [ 24 ] are implemented by ourselves, while the results of other models are taken from [ 16 ] .

Methods	00028			00034			00087
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
X-Avatar [ 29 ]	28.57	0.976	0.026	28.05	0.965	0.035	30.89	0.970	0.030
ExAvatar [ 24 ]	30.58	0.981	0.018	28.75	0.966	0.029	32.01	0.972	0.025
SFGS (ours)	31.12	0.983	0.018	29.10	0.967	0.035	32.10	0.972	0.025

Table 3: Quantitative comparison on subjects 00028 , 00034 , and 00087 on the X-Humans dataset . Higher PSNR/SSIM and lower LPIPS indicate better performance.

Subset	Method	CD $\downarrow$			CD-MAX $\downarrow$			NC $\uparrow$			IoU $\uparrow$
Subset	Method	All	Face	Hand	All	Face	Hand	All	Face	Hand	All	Face	Hand
00028	ExAvatar	9.86	6.30	7.30	82.13	22.29	22.43	0.916	0.837	0.932	0.924	0.751	0.854
00028	Ours	9.06	5.88	7.11	79.39	22.77	22.04	0.927	0.831	0.934	0.939	0.785	0.870
00034	ExAvatar	9.34	5.32	6.93	68.58	21.22	26.03	0.900	0.823	0.941	0.946	0.771	0.905
00034	Ours	9.12	5.60	6.46	67.91	20.47	26.62	0.912	0.825	0.965	0.945	0.825	0.923
00087	ExAvatar	8.85	4.43	6.18	66.64	19.90	23.65	0.886	0.819	0.931	0.949	0.746	0.901

Table 4: More detailed quantitative results. The results verify that our method can produce more detailed and high-quality human reconstructions.

Figure 4: Quality comparison on the Neuman dataset. For each example numbered from (a) to (f), the results are shown from left to right as Ground Truth, SFGS (ours), and ExAvatar [ 24 ] .

Figure 5: Qualitative Comparison on X-humans. From left to right: Ground Truth,Ours and ExAvatar.

Figure 6: Visualizations of the effect of fine-grained hand reconstruction. The SMPL-X mesh shows noticeable artifacts in the hands, which are mitigated by our approach.

Figure 7: Visualizations of the effect of color offset in the structure-aware offset prediction module. The color offset makes the appearance of wrinkles and skin more consistent with real photos. For each approach, the PSNR heat map is on the right.

Figure 8: Visualizations of the effect of structure offset in the structure-aware offset prediction module. The structure offset leads to more accurate mouth pose details.

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Ours w/o Color Offset	30.83	0.983	0.019
Ours w/o Hand Recon.	30.92	0.927	0.019
Ours w/o HexPlane	31.03	0.983	0.019
Ours w/o Structure Offset	30.58	0.982	0.02
SFGS (ours)	31.12	0.983	0.018

Table 5: Results of ablation studies of the proposed SFGS. Removing any single component results in performance drops, especially for the PSNR metric which evaluates fine details.

Figure 9: Temporal consistency comparison with and without HexPlane. The y-axis denotes the tc-LPIPS metric evaluated on 7 consecutive frames. Models equipped with HexPlane exhibit consistently lower LPIPS values, suggesting enhanced temporal coherence.

Hyperparameters	PSNR ↑	SSIM ↑	LPIPS ↓
$\lambda_{\mathrm{a}}=0.3$ , $\lambda_{\mathrm{g}}=0.03$	30.83	0.968	0.02
$\lambda_{\mathrm{a}}=0.3$ , $\lambda_{\mathrm{g}}=0.005$	30.95	0.976	0.19
$\lambda_{\mathrm{a}}=0.5$ , $\lambda_{\mathrm{g}}=0.01$	30.65	0.982	0.02
$\lambda_{\mathrm{a}}=0.1$ , $\lambda_{\mathrm{g}}=0.01$	30.93	0.975	0.02
$\lambda_{\mathrm{a}}=0.3$ , $\lambda_{\mathrm{g}}=0.01$	31.12	0.983	0.018

Table 6: Sensitivity analysis of hyperparameters.

Method	FPS $\uparrow$
Neuman	0.02
Vid2avatar	0.04
X-avatar	0.12
Exavatar	26.00
SFGS (Ours)	30.00

Table 7: Comparison of rendering speed (FPS) with state-of-the-art 3D human avatar generation methods.

3D Hand Reconstruction

See all 36 papers

FlexiAvatar: Unified 3D Gaussian Human Avatars Under Arbitrary Body Visibility
Jul 21, 2026Yihalem Yimolal Tiruneh, Muhammad Salman Ali, Uyoung Jeong +53D Gaussian Splatting 3D Reconstruction
UNIST, South Korea · University of Aberdeen, UK · University College London, UK +1
F3G-Avatar : Face Focused Full-body Gaussian Avatar
Apr 10, 2026Willem Menu, Erkut Akdag, Pedro Quesado +23D Avatar Generation 3D Human Reconstruction
AIMS Group, Department of Electrical Engineering, Eindhoven University of Technology
SplitAvatar: One-shot Head Avatar with Autoregressive Gaussian Splitting
May 25, 2026Hongzhe Liao, Chuhua Xian, Hongmin Cai +23D Gaussian Splatting Adaptive Gaussian Densification
South China University of Technology · University of Tokyo · The Hong Kong University of Science and Technology

Jul 21, 2026J/K move · Enter open · S save

FlexiAvatar: Unified 3D Gaussian Human Avatars Under Arbitrary Body Visibility

Yihalem Yimolal Tiruneh, Muhammad Salman Ali, Uyoung Jeong, Muneeb A. Khan+4

UNIST, South Korea · University of Aberdeen, UK · University College London, UK+1

3D Gaussian Splatting 3D Reconstruction Monocular 3D Reconstruction 3D Avatar Generation 3D Human Reconstruction

Open paper AskPDF arXiv

Reconstructing animatable 3D human avatars from monocular video is a fundamental problem in computer vision with broad applications in AR/VR and digital content creation. Existing approaches typically couple parametric body models with neural rendering or 3D Gaussian splatting and optimize all body regions jointly from short videos, which often degrades fidelity in the visible areas. To overcome this limitation, we introduce FlexiAvatar, a unified framework that explicitly optimizes only the visible body regions, effectively eliminating artifacts arising from unobserved limbs. Our method integrates occlusion-robust SMPL-X tracking with part-specific residual refinement to capture high-frequency geometric and appearance details. To complete entirely unseen regions (e.g., back views), we leverage a diffusion-based approach to generate texture consistent with the observed appearance. Experiments on full-body (NeuMan, ZJU-MoCap, WildAvatar), upper/half-body (talk-show clips), and head-only (INSTA) inputs show that FlexiAvatar delivers consistently higher reconstruction quality, outperforming state-of-the-art methods by an average PSNR improvement of approximately 3% across datasets. Finally, by restricting optimization to observed regions, our method reduces the effective number of Gaussians that must be optimized and rendered, leading to reduced runtime and memory overhead in partial-visibility scenarios.

Abstract

3D Hand Reconstruction

Explore similar work

FlexiAvatar: Unified 3D Gaussian Human Avatars Under Arbitrary Body Visibility

F3G-Avatar : Face Focused Full-body Gaussian Avatar

SplitAvatar: One-shot Head Avatar with Autoregressive Gaussian Splitting