DipGuava: Disentangling Personalized Gaussian Features for 3D Head Avatars from Monocular Video

Jeonghaeng Lee; Seok Keun Choi; Zhixuan Li; Weisi Lin; Sanghoon Lee

cs.CVMar 30, 2026

DipGuava: Disentangling Personalized Gaussian Features for 3D Head Avatars from Monocular Video

Authors: Jeonghaeng Lee, Seok Keun Choi, Zhixuan Li, Weisi Lin, Sanghoon Lee

Organizations: Yonsei University, Korea · Nanyang Technological University, Singapore

Abstract

While recent 3D head avatar creation methods attempt to animate facial dynamics, they often fail to capture personalized details, limiting realism and expressiveness. To fill this gap, we present DipGuava (Disentangled and Personalized Gaussian UV Avatar), a novel 3D Gaussian head avatar creation method that successfully generates avatars with personalized attributes from monocular video. DipGuava is the first method to explicitly disentangle facial appearance into two complementary components, trained in a structured two-stage pipeline that significantly reduces learning ambiguity and enhances reconstruction fidelity. In the first stage, we learn a stable geometry-driven base appearance that captures global facial structure and coarse expression-dependent variations. In the second stage, the personalized residual details not captured in the first stage are predicted, including high-frequency components and nonlinearly varying features such as wrinkles and subtle skin deformations. These components are fused via dynamic appearance fusion that integrates residual details after deformation, ensuring spatial and semantic alignment. This disentangled design enables DipGuava to generate photorealistic, identity-preserving avatars, consistently outperforming prior methods in both visual quality and quantitativeperformance, as demonstrated in extensive experiments.

Figures & tables

Figure 1: Conceptual comparison with prior approaches. (a) Optimization-based methods fail to capture residual details. (b) Entangled models suffer from learning ambiguity. (c) Our disentangled design separately models base and residual features for faithful reconstruction.

Figure 2: Overview of DipGuava . Stage 1 optimizes a geometry-driven base appearance (overall color and opacity from mesh surface normals). Stage 2 predicts residual features to capture facial details beyond the base appearance. Dynamic appearance fusion combines these residuals with the base appearance and geometric deformations.

Figure 3: Dynamic appearance fusion. To ensure alignment, personalized residual appearance sampled from UV space is combined with the base appearance after the geometric deformation.

Method	L1 $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
PointAvatar	0.016	0.056	0.923	25.21
INSTA	0.016	0.108	0.907	24.73
GaussianAvatars	0.012	0.066	0.944	28.36
FlashAvatar	0.018	0.071	0.918	25.72
SplattingAvatar	0.019	0.096	0.917	25.61
MonoGaussianAvatar	0.014	0.053	0.939	26.37

Table 1: Quantitative comparison with SOTA methods. Best and second-best results highlighted.

Figure 4: Qualitative comparison in self-driven animation. Our method generates outperforming results with details such as wrinkles, eye blinks, and lip movement.

Figure 5: Qualitative comparison in cross-id reenactment. The proposed method preserves both facial structure and appearance with high fidelity, while accurately following subtle expressions in the driving motion.

	Training strategy		Model design			Model component					Full
Metric	Joint	LowRes	OptBase	FixedUV	MLP	only $\mathcal{B}$	only $\mathcal{R}$	w/o $\Delta\mathcal{G}$	w/o $\mathcal{R}$	w/o DAF	$\mathcal{B}$ + $\mathcal{R}$ + $\Delta\mathcal{G}$
L1 $\downarrow$	0.013	0.011	0.012	0.010	0.010	0.0099	0.0130	0.0093	0.0092	0.0091	0.0090
LPIPS $\downarrow$	0.0491	0.0949	0.0492	0.0453	0.0516	0.0433	0.0955	0.0440	0.0434	0.0410	0.0417
SSIM $\uparrow$	0.936	0.949	0.944	0.952	0.938	0.9538	0.9357	0.9547	0.9553	0.9554	0.9570
PSNR $\uparrow$	29.02	29.42	28.70	29.53	28.32	29.54	28.25	29.94	29.95	29.86	30.14

Table 2: Quantitative result of training strategies, model designs, and ablations. Best and second-best results are highlighted.

Figure 6: Impact of individual components. The full model preserves structure and fine-scale details most effectively.

Figure 7: Applying dynamic appearance fusion yields sharper and less noisy results in regions with rapid color changes, such as inside the mouth.

Figure 8: Expression-aware dynamic wrinkles. Unlike other methods that model wrinkles as static color textures, our method adaptively captures them in response to facial motion and geometry.

Figure 9: Personalized residuals from the same driving motion. Residual features inject subject-specific details, enabling personalized expressions from identical motions.

3D Avatar Generation

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

See all 90 papers

One-Shot Feed-Forward 360 $^{\circ}$ Animatable Avatar via Inpainted UV-Space Gaussian Modeling
Jan 19, 2026Shuling Zhao, Dan Xu3D Gaussian Splatting 3D Reconstruction
The Hong Kong University of Science and Technology · Zeekr Automobile R&D Co., Ltd.
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
SEGA: Drivable 3D Gaussian Head Avatar from a Single Image
Apr 19, 2025Chen Guo, Zhuo Su, Liao Wang +83D Gaussian Splatting 3D Asset Generation
Tsinghua Shenzhen International Graduate School, China · ByteDance and Pengcheng Lab, China · ByteDance, China +1

Jan 19, 2026J/K move · Enter open · S save

One-Shot Feed-Forward 360 $^{\circ}$ Animatable Avatar via Inpainted UV-Space Gaussian Modeling

Shuling Zhao, Dan Xu

The Hong Kong University of Science and Technology · Zeekr Automobile R&D Co., Ltd.

3D Gaussian Splatting 3D Reconstruction 3D Avatar Generation Human Avatar Animation 3D Face Reconstruction Single-View 3D Reconstruction+1

Open paper AskPDF arXiv

Building one-shot 3D animatable head avatars is an important yet challenging problem. Existing methods generally collapse under large camera pose variations, compromising the realism of 3D avatars. In this work, we propose a new framework to tackle the novel setting of one-shot 3D full-head animatable avatar reconstruction in a single forward pass via inpainted UV-space Gaussian modeling, enabling 360

^\circ

rendering views and real-time animation. To facilitate efficient animation control, we model 3D head avatars with Gaussian primitives embedded on the surface of a parametric face model within the UV space, and project the input image features to the UV space, resulting in incomplete local UV feature maps. To inpaint the missing regions, we obtain knowledge of full-head geometry and textures from rich 3D full-head priors within a pretrained 3D generative adversarial network (GAN) for global full-head feature extraction and multi-view supervision. Specifically, to enhance the fidelity of 3D reconstruction during inpainting, we take advantage of the symmetric nature of the UV space and human faces to fuse incomplete yet detailed local UV feature maps with the extracted global full-head textures, resulting in inpainted UV Gaussian attribute maps for avatar modeling. Extensive experiments demonstrate that our method is the first to achieve high-quality 3D full-head animatable avatar modeling, significantly improving side and back views while outperforming state-of-the-art animation approaches, thereby improving the realism of 3D animatable avatars.

Abstract

3D Avatar Generation

Explore similar work

One-Shot Feed-Forward 360∘^{\circ}∘ Animatable Avatar via Inpainted UV-Space Gaussian Modeling

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

SEGA: Drivable 3D Gaussian Head Avatar from a Single Image

One-Shot Feed-Forward 360 $^{\circ}$ Animatable Avatar via Inpainted UV-Space Gaussian Modeling