GaussianAvatar-Editor: Photorealistic Animatable Gaussian Head Avatar Editor

Arxiv | Project Page

Xiangyue Liu, Kunming Luo, Heng Li, Qi Zhang, Yuan Liu, Li Yi, Ping Tan^†

3DV 2025

Abstract

We introduce GaussianAvatar-Editor, an innovative framework for text-driven editing of animatable Gaussian head avatars that can be fully controlled in expression, pose, and viewpoint. Unlike static 3D Gaussian editing, editing animatable 4D Gaussian avatars presents challenges related to motion occlusion and spatial-temporal inconsistency. To address these issues, we propose the Weighted Alpha Blending Equation (WABE). This function enhances the blending weight of visible Gaussians while suppressing the influence on non-visible Gaussians, effectively handling motion occlusion during editing. Furthermore, to improve editing quality and ensure 4D consistency, we incorporate conditional adversarial learning into the editing process. This strategy helps to refine the edited results and maintain consistency throughout the animation. By integrating these methods, our GaussianAvatar-Editor achieves photorealistic and consistent results in animatable 4D Gaussian editing. We conduct comprehensive experiments across various subjects to validate the effectiveness of our proposed techniques, which demonstrates the superiority of our approach over existing methods. More results and code are available at: https://xiangyueliu.github.io/GaussianAvatar-Editor/.

Setup

Environment

• Clone this repo

  git clone https://github.com/Lxiangyue/GaussianAvatar-Editor.git
  cd GaussianAvatar-Editor

• Install dependencies to setup a conda environment:

  conda create --name gsavatareditor -y python=3.10
  conda activate gsavatareditor
  conda install pytorch::pytorch torchvision torchaudio -c pytorch
  pip install -r requirements.txt

  We made our modifications on CUDA code in diff-gaussian-rasterization, so clone our version and install (compulsory):

  cd submodules/diff-gaussian-rasterization && rm -r diff_gaussian_rasterization.egg-info && pip install . && cd ..
  cd nvdiffrast && rm -r nvdiffrast.egg-info && pip install . && cd ..
  cd simple-knn && rm -r simple-knn.egg-info && pip install . && cd ..

Download

Data

In our paper, we use part of the NeRSemble dataset. You can download the pre-processed data from

• data (directly accessible).
• or from the official: OneDrive (request here).

Model

• Download the FLAME model: flame_model.

• Download Original Gaussian Avatars: gs_origin.

• Download Our Trained Edited Gaussian Avatars Models (Optional, only for inference): outputs.

• Path organized as:

  /GaussianAvatar-Editor
      /data
      /flame_model
      /gs_origin
      /outputs

Reproducing Experiments

Novel View Rendering

1. For example, using editing prompt "Turn him into the Tolkien Elf" on 306_EMO1.

• Inference with our trained model:

  python render.py -m outputs/306_EMO1_Elf  --select_camera_id 8
  # You could change the "select_camera_id" to one of the "0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15" to render the different view.

• Train:

  python train.py -s data/cluster/ikarus/sqian/project/dynamic-head-avatars/code/multi-view-head-tracker/export/306_EMO-1_v16_DS2-0.5x_lmkSTAR_teethV3_SMOOTH_offsetS_whiteBg_maskBelowLine  -m outputs/306_EMO1_Elf_train --port 60001 --eval --white_background --bind_to_mesh --iterations 601_000 --interval 500 --start_checkpoint gs_origin/UNION10EMOEXP_306_eval_600k/chkpnt600000.pth  --use_in2n  --text_prompt "Turn him into the Tolkien Elf" --different_data --use_grad_mask --use_gan  

2. For example, using editing prompt "What would the human look like as a bearded man?" on 104_EXP5.

• Inference with our trained model:

  python render.py -m outputs/104_EXP5_bearded  --select_camera_id 8

• Train:

  python train.py -s data/cluster/ikarus/sqian/project/dynamic-head-avatars/code/multi-view-head-tracker/export/104_EXP-5_v16_DS2-0.5x_lmkSTAR_teethV3_SMOOTH_offsetS_whiteBg_maskBelowLine  -m outputs/104_EXP5_bearded_train --port 60000 --eval --white_background --bind_to_mesh --iterations 601_000 --interval 500 --start_checkpoint gs_origin/UNION10EMOEXP_104_eval_600k/chkpnt600000.pth  --use_in2n  --text_prompt "What would she look like as a bearded man?" --different_data --use_grad_mask --use_gan 

Self-reenactment

1. For example, using editing prompt "Make it an Egyptian sculpture" on 264_EXP5.

• Inference with our trained model:

  python render.py -m outputs/264_EXP5_Egyptian  --select_camera_id 8

• Train:

  python train.py -s data/cluster/ikarus/sqian/project/dynamic-head-avatars/code/multi-view-head-tracker/export/264_EXP-5_v16_DS2-0.5x_lmkSTAR_teethV3_SMOOTH_offsetS_whiteBg_maskBelowLine  -m outputs/264_EXP5_Egyptian_train --port 60000 --eval --white_background --bind_to_mesh --iterations 601_000 --interval 500 --start_checkpoint gs_origin/UNION10EMOEXP_264_eval_600k/chkpnt600000.pth  --use_in2n  --text_prompt "Make it an Egyptian sculpture" --different_data --use_grad_mask --use_gan 

2. For example, using editing prompt "The human should look 100 years old" on 304_EXP2.

• Inference with our trained model:

  python render.py -m outputs/304_EXP2_100  --select_camera_id 8

• Train:

  python train.py -s data/cluster/ikarus/sqian/project/dynamic-head-avatars/code/multi-view-head-tracker/export/304_EXP-2_v16_DS2-0.5x_lmkSTAR_teethV3_SMOOTH_offsetS_whiteBg_maskBelowLine  -m outputs/304_EXP2_100_train --port 60000 --eval --white_background --bind_to_mesh --iterations 601_000 --interval 500 --start_checkpoint gs_origin/UNION10EMOEXP_304_eval_600k/chkpnt600000.pth  --use_in2n  --text_prompt "The human should look 100 years old" --different_data --use_grad_mask --use_gan 

3. For example, using editing prompt "Apply face paint" on 304_EXP2.

• Inference with our trained model:

  python render.py -m outputs/304_EXP2_facepaint  --select_camera_id 8

• Train:

  python train.py -s data/cluster/ikarus/sqian/project/dynamic-head-avatars/code/multi-view-head-tracker/export/304_EXP-2_v16_DS2-0.5x_lmkSTAR_teethV3_SMOOTH_offsetS_whiteBg_maskBelowLine  -m outputs/304_EXP2_facepaint_train --port 60000 --eval --white_background --bind_to_mesh --iterations 601_000 --interval 500 --start_checkpoint gs_origin/UNION10EMOEXP_304_eval_600k/chkpnt600000.pth  --use_in2n  --text_prompt "Apply face paint" --different_data --use_grad_mask --use_gan 

Cross-identy Reenactment

• Please note that Cross-identy Reenactment does not require training.

1. For example, using the source actor 460_FREE to drive the trained edited avatars 306_EMO1_Elf to do the same actions.

python render.py -t /home/super/Desktop/8TDisk/Codes_GaussianAvatarEditor/GaussianAvatar-Editor/data/cluster/ikarus/sqian/project/dynamic-head-avatars/code/multi-view-head-tracker/export/460_FREE_v16_DS2-0.5x_lmkSTAR_teethV3_SMOOTH_offsetS_whiteBg_maskBelowLine -m outputs/306_EMO1_Elf --select_camera_id 8 

2. For example, using the source actor 460_FREE to drive the trained edited avatars 304_EXP2_100 to do the same actions.

python render.py -t /home/super/Desktop/8TDisk/Codes_GaussianAvatarEditor/GaussianAvatar-Editor/data/cluster/ikarus/sqian/project/dynamic-head-avatars/code/multi-view-head-tracker/export/460_FREE_v16_DS2-0.5x_lmkSTAR_teethV3_SMOOTH_offsetS_whiteBg_maskBelowLine -m outputs/304_EXP2_100 --select_camera_id 8 

Citation

If you find our work useful in your research, please cite:

@article{liu2025gaussianavatar,
  title={GaussianAvatar-Editor: Photorealistic Animatable Gaussian Head Avatar Editor},
  author={Liu, Xiangyue and Luo, Kunming and Li, Heng and Zhang, Qi and Liu, Yuan and Yi, Li and Tan, Ping},
  journal={arXiv preprint arXiv:2501.09978},
  year={2025}
}

Acknowledgements

The implementation of GaussianAvatar-Editor are based on GaussianAvatars. Thanks to these authors for releasing the code.