EGGHead: Efficient Generalizable Gaussian Heads from a Single Image

Pilseo Park; Andrew Hou; Michel Sarkis; Ning Bi; Yiying Tong; Xiaoming Liu

cs.CVOct 16, 2026

EGGHead: Efficient Generalizable Gaussian Heads from a Single Image

Authors: Pilseo Park, Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, Xiaoming Liu

Organizations: Michigan State University · Qualcomm Technologies, Inc. · University of North Carolina at Chapel Hill

Abstract

Creating photorealistic human heads typically involves trade-offs between representation power, efficiency, and generalizability. Traditional mesh-based solutions relying on 3DMMs generally sacrifice representation power but are efficient to render. On the other hand, NeRF-based methods improve representation power and fidelity while sacrificing efficiency. Gaussian splatting methods attempt to resolve these trade-offs by offering both high representation power and efficiency with real-time rendering. However, their efficiency nonetheless depends on the number of rasterized Gaussians, and the time needed to optimize and prepare the head (the enrollment time) is often high. In light of this, we propose EGGHead, which introduces a novel Gaussian allocation algorithm that efficiently assigns Gaussians to the head's semantic regions given a fixed Gaussian budget. Our algorithm reduces the number of Gaussians needed while minimizing the impact on representation power. EGGHead operates in a single forward pass and outperforms previous single-image face modeling methods in reconstruction quality, novel view synthesis, and rendering speeds.

Figures & tables

Figure 2: Training Pipeline. We propose a single-image, generalizable, and lightweight Gaussian head model that adaptively allocates M Gaussians with minimal impact on rendering quality. Secs. 3.1–3.4 constitute the main reconstruction pipeline: GAM first allocates M Gaussians from the source image, Voronoi tessellation provides sparse geometric priors, and image/depth features are used to predict Gaussian attributes for rendering and refinement. Sec. 3.5 is trained in parallel to predict a density map supervised by the GAM-derived density map. It is independent of the main reconstruction pipeline during training and is used at inference time for density-based Gaussian allocation.

Figure 3: Inference-time Pipeline. During inference, we replace GAM with density-based sampling to reduce computation, and incorporate a Face Symmetry Module to complete occluded regions.

Panel of Figure 4 — Figure 4: Novel View Synthesis. We qualitatively compare our novel view synthesis performance.

Panel of Figure 5 — Figure 5: Reconstruction Quality. We compare the reconstruction quality on the FFHQ and CelebA-HQ datasets. (a) E $3$ DGE; (b) G-NeRF; (c) GAGAvatar; (d) Ours.

Method	FFHQ			CelebA-HQ
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
E $3$ DGE [ 52 ]	$23.43$	$0.598$	$0.133$	$23.64$	$0.589$	$0.128$
G-NeRF [ 98 ]	$19.68$	$0.673$	$0.256$	$19.81$	$0.566$	$0.296$
GAGAvatar [ 17 ]	$25.03$	$0.888$	$0.110$	$25.06$	$0.874$	$0.119$
Ours	$\mathbf{29.75}$	$\mathbf{0.951}$	$\mathbf{0.017}$	$\mathbf{29.25}$	$\mathbf{0.939}$	$\mathbf{0.018}$

Table 1: Reconstruction Quality. We quantitatively compare the reconstruction quality.

Method	FFHQ		CelebA-HQ
Method	ID $\uparrow$	FID $\downarrow$	ID $\uparrow$	FID $\downarrow$
E $3$ DGE [ 52 ]	$0.7744$	$41.60$	$0.7547$	$40.87$
G-NeRF [ 98 ]	$0.6621$	$54.67$	$0.6375$	$50.45$
GAGAvatar [ 17 ]	$0.8377$	$32.55$	$0.8480$	$\mathbf{30.01}$
Ours	$\mathbf{0.8517}$	$\mathbf{29.01}$	$\mathbf{0.8532}$	$32.23$

Table 2: Novel View Synthesis. We compare novel view synthesis quality using pairwise multi-view identity consistency (ID) and Frechet Inception Distance (FID).

Panel of Figure 6 — Figure 6: NeRSemble NVS. (a) E $3$ DGE; (b) G-NeRF; (c) GAGAvatar; (d) Ours.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
E $3$ DGE [ 52 ]	$20.60$	$0.777$	$0.136$
G-NeRF [ 98 ]	$19.26$	$0.768$	$0.135$
GAGAvatar [ 17 ]	$21.43$	$0.832$	$0.101$
Ours	$\mathbf{22.91}$	$\mathbf{0.835}$	$\mathbf{0.096}$

Table 3: NeRSemble NVS. Quantitative comparison on the NeRSemble dataset.

Method	Enrollment time (sec)	Render speed (FPS)
E $3$ DGE [ 52 ]	$8.63$	$5.26$
G-NeRF [ 98 ]	$5.96$	$33.61$
GAGAvatar [ 17 ]	$\mathbf{0.06}$	$43.43$
Ours	$1.54$	$\mathbf{121.19}$

Table 4: Efficiency. We compare efficiency in enrollment time and rendering speed.

Panel of Figure 7 — Figure 7: Ablation on Sampling Strategy. $M{=}12$ k . (a) Landmark; (b) Parsing; (c) Uniform; (d) Ours.

$M$	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
$12$ k	Landmark	28.64	0.936	0.024
	Parsing	29.02	0.943	0.022
	Uniform	29.13	0.953	0.019
	Adaptive (Ours)	29.75	0.951	0.017
$8$ k	Landmark	28.06	0.918	0.032
	Parsing	28.35	0.925	0.030

Table 5: Ablation on Sampling Strategy. We evaluate whether our allocation strategy (adaptive) more effectively distributes Gaussians under a fixed budget

M

compared to uniform and landmark sampling. Our method achieves better reconstruction quality across

M

settings, with lower SSIM than uniform sampling.

Panel of Figure 8 — Figure 8: Ablation on Voronoi Tessellation. We show the impact of Voronoi tessellation (VT) on capturing 3D Gaussian shapes under single-view and sparse-Gaussian settings.

Panel of Figure 9 — Figure 9: Ablation on Mirroring Gaussians. We show the effect of mirroring Gaussians for handling occluded regions. Our method naturally fills in the occluded areas.

Layer	Filter Size	Input Channel	Output Channel	Activation
Conv	$5\times 5$	$3$	$32$	ReLU
$\mathcal{E}_{1}$	$3\times 3$	$32$	$32$	ReLU
$\mathcal{E}_{2}$	$3\times 3$	$32$	$48$	ReLU
$\mathcal{E}_{3}$	$3\times 3$	$48$	$96$	ReLU
$\mathcal{D}_{3}$	$3\times 3$	$192$	$96$	ReLU
$\mathcal{D}_{2}$	$3\times 3$	$192$	$64$	ReLU

Table 6: Structure of Gaussian Attribute Regressor.

Layer	Filter Size	Input channel	Output channel	Activation
$\mathcal{E}_{1}$	$3\times 3$	$3$	$32$	ReLU
$\mathcal{E}_{2}$	$3\times 3$	$32$	$64$	ReLU
$\mathcal{E}_{3}$	$3\times 3$	$64$	$128$	ReLU
$\mathcal{E}_{4}$	$3\times 3$	$128$	$256$	ReLU
$\mathcal{D}_{3}$	$2\times 2$	$256$	$128$	ReLU
$\mathcal{D}_{2}$	$2\times 2$	$128$	$64$	ReLU

Table 7: Structure of Density Prediction Network.

Layer	Filter Size	Input channel	Output channel	Activation
$\mathcal{E}_{1}$	$3\times 3$	$4$	$48$	SiLU
$\mathcal{E}_{2}$	$3\times 3$	$48$	$96$	SiLU
$\mathcal{E}_{3}$	$3\times 3$	$96$	$192$	SiLU
$\mathcal{E}_{4}$	$3\times 3$	$192$	$384$	SiLU
$\mathcal{M}$	$3\times 3$	$384$	$384$	SiLU
$\mathcal{D}_{3}$	$3\times 3$	$576$	$192$	SiLU

Table 8: Structure of Refinement Network.

Panel of Figure 10 — Figure 10: Edge-aware alignment. Aligning Gaussian positions along the edges improves boundary consistency and reduces artifacts compared to the unaligned configuration.

Panel of Figure 11 — Figure 11: Predicted Density Map. The predicted density maps align well with the GT density maps and assign higher density to fine detail regions which enables adaptive Gaussian allocation.

Number of Gaussians ( M )	Rendering speed (FPS)
$120$ k	179.91
$80$ k	214.01
$40$ k	240.24
$12$ k	309.60
$8$ k	316.62
$4$ k	331.11

Table 9: Rendering Speed. We report rendering speed for different numbers of Gaussians without additional refinement steps.

Panel of Figure 12 — Figure 12: Expression Editing. We show the expression from the driving image is properly transferred to the source identity by deforming Gaussians bound to the mesh.

Panel of Figure 13 — Figure 13: Additional Comparisons. We show improvements in preserving facial details, hair texture, and eye structure/color, and also better reconstructing occluded face regions ( e.g. recovering the subject’s occluded ear in the fourth row).

$M$	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
24 $k$	GAGAvatar [ 17 ]	$21.72$	$\mathbf{0.837}$	$0.100$
12 $k$	Ours	$\mathbf{22.91}$	$0.835$	$\mathbf{0.096}$

Table 10: Additional NeRSemble NVS. Additional quantitative comparison on the NeRSemble dataset using 12

k

Gaussians for our method and 24

k

Gaussians for GAGAvatar.

Panel of Figure 14 — Figure 14: Delighting Preprocessing. We apply an off-the-shelf delighting method to the input image during enrollment to reduce asymmetric illumination artifacts.

3D Face Reconstruction

3D face reconstruction estimates facial shape and appearance from images or video to create usable models and avatars.

See all 54 papers

Large-Scale High-Quality 3D Gaussian Head Reconstruction from Multi-View Captures
May 5, 2026Evangelos Ntavelis, Sean Wu, Mohamad Shahbazi +213D Gaussian Splatting Multi-View 3D Reconstruction
Apple
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
Split and Drive: Dual-Axis Disentanglement for Real-Time Gaussian Head Avatars
Jul 30, 2026MD Wahiduzzaman Khan, Mingshan Jia, Xiaolin Zhang +23D Gaussian Splatting Disentangled Representation Learning
University of Technology Sydney, Australia · Shandong University of Science and Technology, China

May 5, 2026J/K move · Enter open · S save

Large-Scale High-Quality 3D Gaussian Head Reconstruction from Multi-View Captures

Evangelos Ntavelis, Sean Wu, Mohamad Shahbazi, Fabio Maninchedda+20

Apple

3D Gaussian Splatting Multi-View 3D Reconstruction 3D Human Reconstruction 3D Face Reconstruction Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

We propose HeadsUp, a scalable feed-forward method for reconstructing high-quality 3D Gaussian heads from large-scale multi-camera setups. Our method employs an efficient encoder-decoder architecture that compresses input views into a compact latent representation. This latent representation is then decoded into a set of UV-parameterized 3D Gaussians anchored to a neutral head template. This UV representation decouples the number of 3D Gaussians from the number and resolution of input images, enabling training with many high-resolution input views. We train and evaluate our model on an internal dataset with more than 10,000 subjects, which is an order of magnitude larger than existing multi-view human head datasets. HeadsUp achieves state-of-the-art reconstruction quality and generalizes to novel identities without test-time optimization. We extensively analyze the scaling behavior of our model across identities, views, and model capacity, revealing practical insights for quality-compute trade-offs. Finally, we highlight the strength of our latent space by showcasing two downstream applications: generating novel 3D identities and animating the 3D heads with expression blendshapes.

Abstract

3D Face Reconstruction

Explore similar work

Large-Scale High-Quality 3D Gaussian Head Reconstruction from Multi-View Captures

SplitAvatar: One-shot Head Avatar with Autoregressive Gaussian Splitting

Split and Drive: Dual-Axis Disentanglement for Real-Time Gaussian Head Avatars