ELITE: Efficient Gaussian Head Avatar from a Monocular Video via Learned Initialization and TEst-time Generative Adaptation

Kim Youwang · Lee Hyoseok · Park Subin · Gerard Pons-Moll · Tae-Hyun Oh

CVPR 2026

Official implementation of the CVPR 2026 paper, "ELITE: Efficient Gaussian Head Avatar from a Monocular Video via Learned Initialization and TEst-time Generative Adaptation".

What's ELITE?

• ELITE is an efficient system for synthesizing high-fidelity, animatable Gaussian avatars from a short monocular video.
• ELITE leverages a mutually reinforcing synergy of 2D & 3D face priors (generative & data priors) to tackle the longstanding challenge in monocular synthesis of animatable, photorealistic 3D face avatar — Balancing between in-the-wild generalization and efficient synthesis.

Environment Setup

Tested on Ubuntu 24.04, NVIDIA RTX A6000 (48GB), CUDA-11.8, gcc-11, g++-11. Later versions should work, but haven't tested.

git clone git@github.com:kaist-ami/ELITE.git
cd ELITE
git submodule update --init --recursive
export CUDA_HOME=/usr/local/cuda-11.8
export CC=/usr/bin/gcc-11
export CXX=/usr/bin/g++-11
conda create -n ELITE python=3.10
conda activate ELITE
pip install -r requirements.txt
pip install --no-index --no-cache-dir pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py310_cu118_pyt201/download.html
CC=gcc-11 CXX=g++-11 pip install --no-build-isolation git+https://github.com/hbb1/diff-surfel-rasterization.git
pip install -e ./vhap

Data preparation & Configuration

1. Download ELITE checkpoints

Download ELITE's pre-trained 3D prior model and 2D generative prior model checkpoints from this link. Put the downloaded checkpoints under checkpoints/{2d_prior,3d_prior}.pth.

2. Download FLAME-relaed assets

Download FLAME related assets from FLAME. Before you continue, you must register at https://flame.is.tue.mpg.de/ and agree to the license terms.

• First download FLAME 2023 (revised eye region, improved expressions, versions w/ and w/o jaw rotation) and put it under asset/flame/flame2023.pkl.
• Then, download FLAME Vertex Masks and put it under asset/flame/FLAME_masks.pkl

3. Download INSTA videos

We mainly use monocular videos from INSTA (Zielonka, CVPR'23) dataset and create personalized avatars. Check their instructions and download videos.

4. Configure paths (One-time setup)

Open configs/paths.sh and edit the variables at the top to match your environment — in particular CONDA_INIT_SCRIPT (path to your conda initializer) and CUDA_HOME. Then activate the config:

source configs/paths.sh

ELITE: Build your own 3D animatable head avatar!

ELITE is an end-to-end pipeline to create and render a personalized avatar from a monocular video.

Step 0: Prepare source (input) video

Place the input video (e.g., bala.mp4 from INSTA) at data/source/input_videos/{ID}.mp4. A sample input video can be downloaded from this link.

Note: We currently support two resolutions (heightxwidth): 512x512 and 802x550.

Step 1: Process source video

Runs video preprocessing, face 3DMM tracking, and NeRF/3DGS-format export.

# bash scripts/process_source_video.sh {ID}
bash scripts/process_source_video.sh bala

# if you want to use specific gpu index
bash scripts/process_source_video.sh bala 4

The processed artifacts are saved at:

# pre-processed video
data/source/input_videos/{ID}

# 3DMM tracked dataset
data/source/tracked/{ID}_whiteBg_staticOffset

# post-processed data in NeRF/3DGS-format
data/source/processed/{ID}_whiteBg_staticOffset_maskBelowLine  

Step 2: Personalize avatar

Personalize the generalized 3D prior model to the target identity via ELITE's test-time generative adaptation.

# bash scripts/personalize.sh {ID}
bash scripts/personalize.sh bala

# if you want to use specific gpu index
bash scripts/personalize.sh bala 4

Step 3: Render animated face avatars

Renders the personalized avatar driven by a motion sequence. Please download some preset driver motion sequences from this link, and put them under data/drive.

# bash scripts/render_videos.sh {ID} {MOTION_NAME}
bash scripts/render_videos.sh bala nersemble_416_EMO-1-shout+laugh

# if you want to use specific gpu index or FPS
bash scripts/render_videos.sh bala nersemble_416_EMO-1-shout+laugh 4 30

If you want to use your own motion sequence for a driving signal, you can run the preprocessing scripts from the Step 1 (scripts/process_source_video.sh) and put the "processed" data directories under data/drive.

The rendered output videos are saved at:

outputs/{ID}/vis_motion/{ID}_rgb_{MOTION_NAME}.mp4   # RGB render
outputs/{ID}/vis_motion/{ID}_nrm_{MOTION_NAME}.mp4   # Normal map render

Citation

If you find our code or paper helps, please consider citing:

@inproceedings{youwang2026elite,
    title = {ELITE: Efficient Gaussian Head Avatar from a Monocular Video via Learned Initialization and TEst-time Generative Adaptation},
    author = {Youwang, Kim and Hyoseok, Lee and Subin, Park and Pons-Moll, Gerard and Oh, Tae-Hyun},
    booktitle = {CVPR},
    year = {2026}
}

Contact

Kim Youwang (youwang.kim@postech.ac.kr)