SVAD: From Single Image to 3D Avatar via Synthetic Data Generation with Video Diffusion and Data Augmentation
SDAV creates photorealistic 3D human avatars from a single image input. Our approach combines video diffusion models with 3D Gaussian Splatting to generate high-fidelity, animatable avatars while preserving the subject's identity and appearance details.
Key Features:
- Single-image input to full-body 3D avatar generation
- [2025/05/08] Code has been released.
- [2025/03.30] SVAD has been accepted to CVPR 2025 Workshop.
- CUDA 12.1
- Python 3.9
- PyTorch 2.3.1
First you need to clone the repo:
git clone https://github.com/yc4ny/SVAD.git
cd SVAD
Our default installation method is based on Conda package and environment management:
cd envs/
conda env create --file svad.yaml
conda activate svadInstall required packages via pip:
pip install pip==24.0 # For pytorch-lightning==1.6.5
pip install cmake # For dlib
pip install -r svad.txt Install CUDA related packages:
conda install -c "nvidia/label/cuda-12.1.1" cuda-nvcc cuda-cudart-dev libcurand-dev
conda install -c nvidia cuda-profiler-apiInstall MMPose:
pip install --no-cache-dir -U openmim
pip install mmdet
pip install mmcv==2.2.0 -f https://download.openmmlab.com/mmcv/dist/cu121/torch2.3/index.html
pip install mmposeInstall modified version of 3DGS:
cd diff_gaussian_rasterization_depth/
python setup.py install
cd .. # Return to envs directoryInstall 3DGS Renderer:
cd diff-gaussian-rasterization
pip install .
cd ../../ # Return to base avatar directoryYou can download the required files to run this code with:
gdown 1_dH5-0fMTfED_BI1h_fQcL2bHyLqNBot
python utils/extract_zip.py # This will put extracted files in place.
rm -rf Avatar.zipAlternatively, use this google drive link.
To ensure seamless execution of our pipeline, certain package modifications are required. Please apply the fixes detailed in FIX.md before running the code. These modifications address compatibility issues with dependencies and will prevent runtime errors during execution.
We provide a script to automate the entire pipeline for avatar generation from a single image. This script supports parallel processing (training multiple avatars at the same time 1 per 1 gpu).
Put your input image in data/{name_of_image}/ref.png.
In our example we place image in data/female-3-casual/ref.png/.
# For a single avatar
bash script.sh <name_of_image> <gpu_id>
# Examples for parallel processing on multiple GPUs
bash script.sh female-3-casual 0 # Running on GPU 0
bash script.sh thuman_181 1 # Running on GPU 1If you have correctly installed the environment and applied the package fixes in FIX.md, there should be no problem running the automated script. However, if there are issues, we also provide a step by step guide to train your avatar. Follow the details in TRAIN.md.
After training your avatar, you can render the rotating avatar with neutral pose:
cd avatar/main/
python get_neutral_pose.py --subject_id {SUBJECT_ID} --test_epoch 4You can animate your avatar with :
python animation.py --subject_id {SUBJECT_ID} --test_epoch 4 --motion_path {MOTION_PATH}The rendering code is based on ExAvatar so to prepare motion sequences, please refer to this link.
Parts of the code are taken or adapted from the following repos:
If you find this code useful for your research, please consider citing the following paper:
@inproceedings{choi2025svad,
title={SVAD: From Single Image to 3D Avatar via Synthetic Data Generation with Video Diffusion and Data Augmentation},
author={Yonwoo Choi},
booktitle={CVPR 2025 Workshop SyntaGen: 2nd Workshop on Harnessing Generative Models for Synthetic Visual Datasets},
year={2025}
}