F3G-Avatar

F3G-Avatar: Face Focused Full-body Gaussian Avatar · CVPRW 2026 · Paper · Code · Project Page

Official PyTorch implementation for reconstructing realistic, animatable full-body human avatars from multi-view RGB video.

Introduction

F3G-Avatar is a face-aware full-body avatar synthesis method. Starting from a clothed Momentum Human Rig (MHR) template, it renders front/back positional maps decoded into 3D Gaussians through a two-branch architecture: a body branch for pose-dependent non-rigid deformations and a face-focused branch for facial geometry and appearance. Gaussians are fused, posed with linear blend skinning (LBS), and rendered with differentiable Gaussian splatting. Training combines reconstruction and perceptual losses with a face-specific adversarial loss.

F3G-Avatar displays state-of-the-art rendering quality by delivering improved facial details.

Models

Checkpoint	Description	Dataset	Status
avatarrex_zzr	Body + face trained avatar	AvatarReX	Coming soon

Pretrained weights will be uploaded here under checkpoints/. Until then, train from scratch using the GitHub repository.

Download (when available)

from huggingface_hub import hf_hub_download

ckpt = hf_hub_download(
    repo_id="wjmenu/F3G-Avatar",
    filename="checkpoints/avatarrex_zzr/epoch_latest.pt",
)

Quick start

Code, installation, data preparation, and training live on GitHub (kept in one place to avoid duplication):

git clone https://github.com/wjmenu/F3G-avatar.git
cd F3G-avatar
conda create -n animatable_gaussians python=3.10 -y
conda activate animatable_gaussians
pip install -r requirements.txt
# See README for CUDA extensions, SMPL-X, NeuS2, and MHR template pipeline

Train:

python main_avatar.py -c configs/avatarrex_zzr/avatar.yaml -m train

Citation

@misc{menu2026f3gavatarfacefocused,
  title={F3G-Avatar : Face Focused Full-body Gaussian Avatar},
  author={Willem Menu and Erkut Akdag and Pedro Quesado and Yasaman Kashefbahrami and Egor Bondarev},
  year={2026},
  eprint={2604.09835},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2604.09835},
}

Acknowledgements

Built on Animatable Gaussians, NeuS2, 4D-Dress, PhysAvatar, and StyleAvatar.