Generalizable and Animatable 3D Full-Head Gaussian Avatar from a Single Image

TL;DR

This work tackles one-shot 3D full-head avatar reconstruction from a single image by binding Gaussian primitives to the FLAME surface in UV space and leveraging global full-head priors from a pretrained 3D GAN via feed-forward inversion. It introduces symmetric UV space feature fusion to combine global full-head information with local input details, and enforces a 3D total-variation loss to ensure complete surface coverage and reduce holes. The approach enables high-fidelity, 360° renderable avatars that can be real-time animated (≈$246$ FPS), addressing the limitations of large pose variations in prior methods and advancing practical 3D talking-head applications. The combination of UV-space Gaussian modeling, 3D GAN priors, and symmetry-aware fusion offers a scalable, one-shot solution for realistic, manipulable 3D head avatars with potential impact on teleconferencing, VR, and AR experiences.

Abstract

Building 3D animatable head avatars from a single image 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 feed-forward pass, enabling real-time animation and simultaneous 360$^\circ$ rendering views. 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. To obtain knowledge of full-head geometry and textures, we leverage rich 3D full-head priors within a pretrained 3D generative adversarial network (GAN) for global full-head feature extraction and multi-view supervision. To increase the fidelity of the 3D reconstruction of the input image, we take advantage of the symmetric nature of the UV space and human faces to fuse local fine-grained input image features with the global full-head textures. Extensive experiments demonstrate the effectiveness of our method, achieving high-quality 3D full-head modeling as well as real-time animation, thereby improving the realism of 3D talking avatars.

Figures (9)

• Figure 1: Overview of the framework. Given an input source image, the UV space feature extraction module extracts its global and local UV feature maps for animatable 3D full-head reconstruction. The symmetric UV space feature fusion module takes advantage of the symmetry of human faces and the UV space to combine these UV feature maps. From the predicted UV Gaussian attribute maps, 3D Gaussian primitives are sampled, which can be animated with a parametric face model and rendered given a camera pose.
• Figure 2: Illustration of Symmetric UV Space Feature Fusion at scale $i$. For each feature patch in $\mathbf{F}_{g}^i$, we query two symmetric local windows corresponding to the patch position in $\mathbf{F}_{l}^i$. The output feature $\mathbf{F}_{c}^i$ is further enhanced by the local UV feature map and its symmetry with convolution.
• Figure 3: Effect of the 3D total variation loss $\mathcal{L}_{3d}$. Compared with the UV total variation loss $\mathcal{L}_{uv}$ from kirschstein2024gghead, $\mathcal{L}_{3d}$ can alleviate holes on the avatar surfaces without bringing additional artifacts. Blue boxes indicate the erroneous areas in the rendered images.
• Figure 4: Multi-view results of our method on the HDTF zhang2021flow dataset. Our avatars can be viewed in 360$^\circ$.
• Figure 5: Qualitative comparison with state-of-the-art methods (i.e., Real3DPortrait ye2024realdportrait, Portrait4D deng2024portrait4d, Portrait4D-v2 deng2024portrait4d_eccv, GAGAvatar chu2024generalizable and LAM he2025lam) for cross-identity reenactment on the VFHQ xie2022vfhq and HDTF zhang2021flow datasets. Our method best maintains source identity while effectively mimicking the driving motion.