DiGS-Avatar: Single-Image Animatable 3D Human Reconstruction via UV-Space Diffusion

Jiakun Li; Li Fang; Hao Zhu; Fei Hu; Long Ye; Yuan Zhang; Jinyao Yan

cs.CVAug 21, 2026

DiGS-Avatar: Single-Image Animatable 3D Human Reconstruction via UV-Space Diffusion

Authors: Jiakun Li, Li Fang, Hao Zhu, Fei Hu, Long Ye, Yuan Zhang, Jinyao Yan

Organizations: Key Laboratory of Media Audio and Video (Communication University of China), Ministry of Education, Beijing 100024, China · School of Intelligence Science and Technology, Nanjing University, Nanjing 210023, China

Abstract

Single-image 3D human reconstruction often suffers from over-smoothed textures and geometric inconsistencies. While diffusion models improve generative quality, their reliance on multi-view synthesis prior to 3D reconstruction is computationally expensive and prone to view inconsistency. We propose DiGS-Avatar, which reformulates this task as an efficient, diffusion-based UV-latent completion task, ensuring 3D consistency by design. To capture accurate spatial structure, we introduce a teacher-student framework where a multi-view teacher provides geometrically aligned pseudo-ground-truth latents to supervise a single-view diffusion student. Treating this inferred latent as a robust structural skeleton, our method injects high-level semantic features to accurately recover fine textural details without disrupting spatial integrity. The refined representation is then decoded into 3D Gaussian primitives. Extensive experiments demonstrate that DiGS-Avatar achieves state-of-the-art or highly competitive visual fidelity and zero-shot generalization, while reconstructing a fully animatable 3D avatar in just 0.71 seconds. Code is available at https://github.com/KLMAV-CUC/DiGS-Avatar.

Figures & tables

Figure 1: Comparison of paradigms for single-image 3D human reconstruction. (a) Feed-forward methods (e.g., IDOL [ 51 ] ) directly regress 3D representations, enable fast inference but struggle with unseen regions and texture diversity. (b) Diffusion-based methods (e.g., Human3Diffusion [ 44 ] ) synthesize multi-view images before reconstruction, incurring high cost and view inconsistency. (c) Our DiGS-Avatar performs diffusion-based UV-latent completion in a canonical, surface-aligned space, achieving both photorealistic and geometrically consistent 3D avatars with efficient inference.

Figure 2: Overview of the proposed DiGS-Avatar framework. Given a single input image, the student network extracts VAE and appearance features. The VAE features are mapped to the canonical UV space via SMPL-X-based inverse texture mapping, forming a partial UV latent that is completed by a diffusion model under teacher supervision. The completed UV latent is enriched by GASA with appearance features, producing an aggregated UV feature map that is decoded into 3D Gaussian primitives for rendering and animation. During training, the multi-view teacher network provides geometrically consistent UV latents as supervision for the student.

Methods	HuGe100K [ 51 ]			THuman 2.1 [ 50 ]			2K2K [ 8 ]
Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
SIFU [ 48 ]	18.44	0.921	0.095	16.82	0.926	0.095	18.54	0.931	0.091
Human3Diffusion [ 44 ]	18.09	0.893	0.101	15.67	0.902	0.105	19.29	0.908	0.079
IDOL* [ 51 ]	22.75	0.926	0.084	18.82	0.937	0.077	18.17	0.916	0.092
LHM* [ 27 ]	21.76	0.919	0.065	19.62	0.931	0.072	20.67	0.918	0.073
TRELLIS [ 38 ]	18.24	0.895	0.093	16.52	0.910	0.094	18.72	0.907	0.084

Table 1: Quantitative comparison on in-distribution datasets. “Ours” refers to the diffusion-based student network used during inference. The best result is highlighted in bold and the second best is underlined . * denotes animatable representations.

Figure 3: Qualitative comparison on datasets (top-down): HuGe100K, THuman 2.1, 2K2K, and SIZER. Deterministic baselines over-smooth unseen regions, while multi-view diffusion methods exhibit geometric inconsistencies. In contrast, DiGS-Avatar produces sharp, 3D-consistent details. Notably, on the out-of-distribution SIZER dataset (bottom), our method uniquely preserves the complex geometry of loose clothing.

Methods	SIZER [ 35 ]			Computational Cost
Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Train Time (in GPU hours)	Test Time (s) $\downarrow$
SIFU [ 48 ]	19.30	0.905	0.101	-	30
Human3Diffusion [ 44 ]	19.14	0.904	0.088	$\sim$ 960	28
IDOL* [ 51 ]	21.90	0.922	0.075	$\sim$ 768	2.25
LHM* [ 27 ]	23.45	0.929	0.064	$\sim$ 2,496	5.13
TRELLIS [ 38 ]	19.94	0.916	0.078	$\sim$ 10k	7.43

Table 2: Zero-shot generalization on SIZER and computational efficiency. DiGS-Avatar achieves competitive zero-shot quality while requiring substantially lower inference and training cost.

Figure 4: Zero-shot reconstruction on in-the-wild images from DeepFashion [ 19 ] and the Internet. Using Multi-HMR [ 3 ] for SMPL-X estimation, our model successfully handles complex real-world illumination, diverse poses, and layered casual clothing.

Figure 5: Animation results. Reconstructed avatars are directly driven by LBS and preserve coherent texture attachment under novel poses.

Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Teacher	26.18	0.942	0.037	w/o Teacher	19.03	0.804	0.168
w/o GASA	22.52	0.915	0.084	w/ LDM-UNet	17.39	0.790	0.204
w/o condition	23.26	0.936	0.061	w/o $\mathcal{L}_{\text{diff}}$	15.98	0.785	0.226
Ours (full)	25.32	0.940	0.039

Table 3: Quantitative ablation of key components on HuGe100K. “Ours” refers to the diffusion-based student network used during inference.

Figure 6: Ablation and Robustness Analysis. Left: Qualitative ablation of key components. Right: Robustness to challenging in-the-wild inputs.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1 : Detailed architecture of key modules: (a) UV-Alignment Transformer. (b) GASA module.

Figure S2 : Additional in-the-wild reconstruction results under diverse clothing, poses, and illumination.

Figure S3 : Additional animation results driven by novel pose sequences.

Metric	Geometric Consistency	Texture Clarity	Animation Quality	Overall
Preference	66.6%	60.0%	56.6%	63.3%

Appendix

Table S1 : User Preference Rates (Ours vs. IDOL & LHM). Percentage of times evaluators preferred DiGS-Avatar over the animatable baselines across 50 challenging cases (

N=30

3D Avatar Generation

3D avatar generation creates realistic, controllable digital humans from images, video, text, or other inputs.

See all 89 papers

S-Avatar: Diffusion-Guided Gaussian Head Avatars from a Single Image
Jul 30, 2026Hail Song, Seokhwan Yang, Jiwon Yang +23D Gaussian Splatting 3D Avatar Generation
KAIST UVR Lab · KAIST KI-ITC ARRC
Any3DAvatar: Fast and High-Quality Full-Head 3D Avatar Reconstruction from Single Portrait Image
Apr 15, 2026Yujie Gao, Yao Xiao, Xiangnan Zhu +43D Avatar Generation 3D Reconstruction
Shanghai Jiao Tong University
GenLCA: 3D Diffusion for Full-Body Avatars from In-the-Wild Videos
Apr 8, 2026Yiqian Wu, Rawal Khirodkar, Egor Zakharov +63D Avatar Generation 3D Diffusion Models
State Key Laboratory of CAD&CG, Zhejiang University, China · Codec Avatars Lab, Meta, USA · State Key Lab of CAD&CG, Zhejiang University, China

Jul 30, 2026J/K move · Enter open · S save

S-Avatar: Diffusion-Guided Gaussian Head Avatars from a Single Image

Hail Song, Seokhwan Yang, Jiwon Yang, Woojin Cho+1

KAIST UVR Lab · KAIST KI-ITC ARRC

3D Gaussian Splatting 3D Avatar Generation Single-View 3D Reconstruction

Open paper AskPDF arXiv

We propose S-Avatar, a novel method for generating photorealistic 3D head avatars from a single image using a diffusion-guided 3D model generation module and strategies for animating 3D Gaussian Splatting (3DGS). While single-image head avatar reconstruction is crucial for lifelike Virtual Reality (VR) applications, existing approaches often struggle to preserve 3D consistency under unseen viewpoints. S-Avatar addresses this limitation through a three-stage pipeline. First, a high-resolution 3DGS is synthesized directly from a single image using a diffusion-based Gaussian splat generation module. Next, the parametric head model FLAME is aligned with the generated 3DGS by optimizing its parameters and spatial transformations. Finally, to adapt the 3DGS to FLAME variations, we construct a binding template that encodes the spatial relationship between the initial splats and FLAME. The dynamic 3D head avatar can then be rendered in real time by deforming the 3DGS with the binding template. By combining diffusion-guided canonical 3DGS generation with FLAME-based control, our method achieves efficient and accurate reconstruction with enhanced 3D consistency. Evaluations on public datasets demonstrate that S-Avatar outperforms state-of-the-art methods in novel-view and expression generation, achieving superior realism and consistency. Consequently, our approach represents a significant advance in accessible avatar creation, applicable to a wide range of VR/AR applications. The project page is available at https://github.com/hailsong/savatar.

Abstract

3D Avatar Generation

Explore similar work

S-Avatar: Diffusion-Guided Gaussian Head Avatars from a Single Image

Any3DAvatar: Fast and High-Quality Full-Head 3D Avatar Reconstruction from Single Portrait Image

GenLCA: 3D Diffusion for Full-Body Avatars from In-the-Wild Videos