EGGHead: Efficient Generalizable Gaussian Heads from a Single Image

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

Submitted: 16 October 2026

Subject: Computer Vision and Pattern Recognition (cs.CV)

arXiv: 2610.26898 [cs.CV]

Abstract

Creating photorealistic human heads typically involves trade-offs between representation power, efficiency, and generalizability. Traditional mesh-based solutions relying on 3D Morphable Models (3DMMs) generally sacrifice representation power but are efficient to render. NeRF-based methods improve representation power and fidelity at the cost of efficiency. Gaussian splatting methods attempt to resolve these trade-offs by offering both high representation power and real-time rendering. However, their efficiency still depends on the number of rasterized Gaussians, and the time required 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 under 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 speed.

Key Contributions

• Semantic-aware Gaussian allocation: A novel algorithm distributes a fixed Gaussian budget across the head's semantic regions, cutting the required Gaussian count without significantly sacrificing representational fidelity.
• Single forward pass: EGGHead generates a photorealistic, generalizable 3D head from a single image in one pass, avoiding lengthy per-subject optimization.
• Balanced trade-offs: The method simultaneously targets high representation power, rendering efficiency, and generalizability—areas where 3DMM-based, NeRF-based, and prior Gaussian splatting approaches typically compromise.
• State-of-the-art results: EGGHead outperforms prior single-image face modeling methods in reconstruction quality, novel view synthesis, and rendering speed.

Context

As 3D Gaussian splatting has matured through 2025–2026 into a mainstream real-time neural rendering technique, attention has shifted from raw visual fidelity toward efficiency, generalization, and deployability. Methods that can reconstruct animatable, photorealistic human heads from minimal input—ideally a single image—are increasingly relevant for telepresence, virtual avatars, digital humans, and AR/VR applications. EGGHead addresses this direction by combining a fixed Gaussian budget with semantic region assignment, reducing both rasterization cost and enrollment time while retaining the representational strengths of Gaussian splatting.

Metadata

• Cite as: arXiv:2610.26898 [cs.CV]
• DOI: https://doi.org/10.48550/arXiv.2610.26898
• Submission history: [v1] 16 Oct 2026

via ArXiv CV