FastAvatar: Towards Unified Fast High-Fidelity 3D Avatar Reconstruction with Large Gaussian Reconstruction Transformers

Yue Wu, Xuanhong Chen*, Yufan Wu, Wen Li, Yuxi Lu, Kairui Feng

Tongji University; Shanghai Innovation Institute; Shanghai Jiao Tong University; AKool

* Corresponding Author

🔥 Overview

FastAvatar is a feedforward 3D avatar framework capable of flexibly leveraging diverse daily recordings (e.g., a single image, multi-view observations, or monocular video) to reconstruct a high-quality 3D Gaussian Splatting (3DGS) model within seconds, using only a single unified model.

🎉 Core Highlights

• Unified Model with Flexible Multi-frame Aggregation for Ultra-high-fidelity Avatars
• Fast Feedforward Framework Delivering High-quality 3D Gaussian Splatting Models in Seconds
• Incremental 3D Avatar Reconstruction Leveraging Diverse Inputs (single-shot, monocular and multi-view)

📹 Demo

Self-reenacted

Cross-reenacted

Multi-view & Incremental Reconstruction

🚀 Get Started

Environment Setup

git clone https://github.com/TyrionWuYue/FastAvatar.git
cd FastAvatar

conda create --name fastavatar python=3.10
conda activate fastavatar

bash ./scripts/install/install_cu124.sh

HuggingFace Download

# Download Assets
# export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download TyrionWuY/FastAvatar-assets --local-dir ./tmp
tar -xf ./tmp/assets.tar 
tar -xf ./tmp/model_zoo.tar
rm -r ./tmp/
# Download Model Weihts
huggingface-cli download TyrionWuY/FastAvatar --local-dir ./model_zoo/fastavatar/

Inference

bash scripts/infer/infer.sh ${CONFIG} ${MODEL_NAME} ${IMAGE_INPUT} ${MOTION_SEQS_DIR} ${INFERENCE_N_FRAMES} ${MODE}

IMAGE_INPUT can be either a .mp4 video file or a folder path containing arbitrary number of images. INFERENCE_N_FRAMES is used to control the number of frames input to the model. MODE has two options: 'Monocular' and 'MultiView'.

Acknowledgement

This work is built on many amazing research works and open-source projects:

• VGGT
• LAM
• GaussianAvatars
• VHAP

Thanks for their excellent works and great contribution.

@misc{wu2025fastavatarunifiedfasthighfidelity,
      title={FastAvatar: Towards Unified Fast High-Fidelity 3D Avatar Reconstruction with Large Gaussian Reconstruction Transformers}, 
      author={Yue Wu and Yufan Wu and Wen Li and Yuxi Lu and Kairui Feng and Xuanhong Chen},
      year={2025},
      eprint={2508.19754},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.19754}, 
}