AvatarPointillist: AutoRegressive 4D Gaussian Avatarization

AvatarPointillist is an autoregressive framework for generating dynamic 4D Gaussian avatars from a single portrait image.

More Research

💡 We also have other avatar projects that may interest you.

HeadArtist: Text-conditioned 3D Head Generation with Self Score Distillation, SIGGRAPH 2024

Follow-Your-Emoji: Fine-Controllable and Expressive Freestyle Portrait Animation, SIGGRAPH Asia 2024

AvatarArtist: Open-Domain 4D Avatarization, CVPR 2025

Overview

AvatarPointillist formulates avatar generation as a sequential prediction problem. Starting from a single portrait image, the method:

• autoregressively generates Gaussian point clouds with a decoder-only Transformer
• jointly predicts per-point binding for animation
• decodes full renderable Gaussian attributes with a dedicated Gaussian decoder
• produces photorealistic and controllable 4D avatars

This repository will host the official implementation once the public release package is ready.

Release Status

The public release materials are currently under internal company review.

At the moment, the following components are not yet available:

• training code
• inference code
• data preprocessing code
• pretrained checkpoints

We will update this repository once the review is complete and the release package is cleared for public use.

Method at a Glance

AvatarPointillist contains two tightly coupled stages:

1. An autoregressive generator predicts Gaussian geometry tokens together with binding information.
2. A Gaussian decoder converts the generated representation into complete renderable Gaussian attributes for animation and rendering.

Conditioning the decoder on latent features from the autoregressive generator substantially improves the final avatar fidelity.

Authors

Hongyu Liu^1,2, Xuan Wang², Zijian Wu², Yating Wang², Ziyu Wan³, Yue Ma¹, Runtao Liu¹, Boyao Zhou², Yujun Shen², Qifeng Chen¹

¹ HKUST, ² Ant Group, ³ City University of Hong Kong

Planned Release

The public release is planned to include the following components:

• environment setup
• checkpoints and pretrained models
• demo and inference scripts
• training pipeline
• data preparation instructions

Citation

If you find this project useful, please cite:

@inproceedings{liu2026avatarpointillist,
  title     = {AvatarPointillist: AutoRegressive 4D Gaussian Avatarization},
  author    = {Hongyu Liu and Xuan Wang and Zijian Wu and Yating Wang and Ziyu Wan and Yue Ma and Runtao Liu and Boyao Zhou and Yujun Shen and Qifeng Chen},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2026}
}

Contact

For project-related questions, please use the author homepages above or refer to the project page:

https://kumapowerliu.github.io/AvatarPointillist/