3DXTalker generates identity-consistent, expressive 3D talking avatars from a single reference image and speech audio, achieving accurate lip synchronization, expressive emotion control, and natural head-pose dynamics. It achieves expressive facial animation through data-curated identity modeling, audio-rich representations, and spatial dynamics controllability. By introducing frame-wise amplitude and emotional cues beyond standard speech embeddings, 3DXTalker delivers superior lip synchronization and nuanced expression modulation. Built on a flow-matching transformer architecture, it enables natural head-pose motion generation while supporting stylized control, integrating lip synchronization, emotional expression, and head-pose dynamics within a unified framework.

• Release the 3DTalking benchmark dataset
  • Release the raw dataset
  • Release the processed dataset
• Release the data processing code
• Release the training and inference code
• Release the pretrained models

• Python 3.10
• Pytorch 2.2.2
• CUDA 12.1
• Pytorch3d 0.7.7

conda create -n env_3DXTalker python==3.10
conda activate env_3DXTalker
pip install -r requirements.txt

For some people the compilation fails during requirements install and works after. Try running the following separately:

pip install "git+https://github.com/facebookresearch/pytorch3d.git@v0.7.7"

1. Download emotion2vec_plus_base model and place it in ./pretrained_models/:

   # Create directory
   mkdir -p pretrained_models/emotion2vec_plus_base
   
   # Option 1: Using git-lfs (recommended)
   cd pretrained_models
   git lfs install
   git clone https://huggingface.co/iic/emotion2vec_plus_base
   
   # Option 2: Manual download from https://huggingface.co/iic/emotion2vec_plus_base
   # Download all files to ./pretrained_models/emotion2vec_plus_base/

2. Download microsoft/wavlm-base-plus (audio encoder):

   # Option 1: Auto-download on first run (recommended)
   # The model will be automatically downloaded from HuggingFace when you run training
   
   # Option 2: Pre-download manually
   cd pretrained_models
   git lfs install
   git clone https://huggingface.co/microsoft/wavlm-base-plus
   
   # Then update config/default_config.yaml:
   # audio_encoder_repo: './pretrained_models/wavlm-base-plus'

   Expected directory structure:

   pretrained_models/
   ├── emotion2vec_plus_base/
   │   ├── config.json
   │   ├── pytorch_model.bin
   │   └── ...
   └── wavlm-base-plus/          # Optional (auto-downloads if not present)
       ├── config.json
       ├── pytorch_model.bin
       └── ...
   

1. Download raw video datasets following these links: V0-GRID; v1-RAVDESS; V2-MEAD; V3-VoxCeleb2; V4-HDTF; V5-Celebv-HQ

   If you don't want to process the data manually, we also provide processed data at Hugging Face.

2. Run data curation (duration, noise, language, sync, resolution normalization).

• Edit raw_video_dir in data_prepare/data_curation_pipeline.py to your raw video folder.

cd data_prepare
python data_curation_pipeline.py

Output will be in data_prepare/final_curated_videos/.

1. Rename videos for dataset indexing.

• Edit dataset_name, input_dir, and output_dir in data_prepare/rename.py if needed.
• By default it expects input at data_prepare/Scaled_videos and outputs to data_prepare/Renamed_videos.

cd data_prepare
python rename.py

1. Download EMOCA-related assets (models and FLAME files).

bash gdl_apps/EMOCA/demos/download_assets.sh

1. Run EMOCA reconstruction to extract FLAME parameters.

• Edit data_root_dir and dataset_name in gdl_apps/EMOCA/demos/my_recons_video.py.
• data_root_dir should contain <dataset_name>/all_videos_path.txt.

python gdl_apps/EMOCA/demos/my_recons_video.py \
  --dataset_name VoxCeleb2 \
  --output_folder video_output \
  --model_name EMOCA_v2_lr_mse_20

1. Data structures are provided in DATASET_STRUCTURE.md

@misc{wang20263dxtalkerunifyingidentitylip,
      title={3DXTalker: Unifying Identity, Lip Sync, Emotion, and Spatial Dynamics in Expressive 3D Talking Avatars}, 
      author={Zhongju Wang and Zhenhong Sun and Beier Wang and Yifu Wang and Daoyi Dong and Huadong Mo and Hongdong Li},
      year={2026},
      eprint={2602.10516},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.10516}, 
}