Tavatar

This repository contains the official implementation of Tavatar (CVPR 2026). A high-quality 3D avatar reconstruction method that leverages 2D image inputs to generate realistic 3D avatars.

News

• [2026-03] 🎉 Tavatar is accepted to CVPR 2026.

Overview

Environment

• Python = 3.9
• CUDA 11.8
• Ubuntu/Linux or Windows with CUDA-enabled GPU
• uv package manager

Install Dependencies

uv sync

Data Preparation

SMPL Models

Download the SMPL model files and place them in the following directory structure:

./dataset/smpl_models/smpl/
  ├── SMPL_FEMALE.pkl
  ├── SMPL_MALE.pkl
  └── SMPL_NEUTRAL.pkl

You can obtain SMPL models from the 🔗official SMPL website (registration required).

Inference

bash scripts/eval.sh # bash/zsh
# or
.\scripts\eval.ps1 # PowerShell

Data Source

• Data preparation follows the scripts provided by 🔗InstantAvatar.
• We extract normal maps using 🔗Sapiens.

Troubleshooting

Windows Build Issues

If you encounter build errors on Windows, ensure:

1. Visual Studio 2022 is installed with "Desktop development with C++" workload
2. CUDA 11.8 is properly installed and in PATH

License

This project is released under the MIT License.

See LICENSE for full license terms.