ACM Transactions on Graphics, 2023

HAvatar: High-fidelity Head Avatar via Facial Model Conditioned Neural Radiance Field

Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu

Tsinghua University

Loading showcase video...

HAvatar synthesizes high-resolution dynamic head avatars with stable animation, controllable pose and expression, and view-consistent rendering.

Abstract

Modeling an animatable 3D human head avatar from a lightweight setup requires both realistic portrait synthesis and accurate expression control. HAvatar introduces a hybrid explicit-implicit representation, Facial Model Conditioned Neural Radiance Field, which integrates parametric facial-model priors with the flexibility of NeRF. Synthetic renderings condition the implicit field without constraining topology, while an image-to-image translation network produces high-resolution, photorealistic, and view-consistent dynamic head appearance.

Method

The representation combines explicit facial-model guidance with an implicit radiance field. This lets the avatar keep stable shape and animation behavior while preserving the flexibility needed for hair, mouth interior, and other details that do not fit a fixed template cleanly.

Results

HAvatar focuses on dynamic head synthesis under controllable pose, expression, and viewpoint. The showcase video highlights self-reenactment and animation behavior, with the local MP4 embedded above for direct playback.

Stable facial animation
Head-pose and expression control
Novel-view rendering
High-resolution photorealistic output

Technical Paper

The paper appears in ACM Transactions on Graphics and details the facial-model conditioned radiance field, the synthetic-rendering conditioning strategy, and the image-to-image translation architecture.

Read the ACM paper View arXiv version

Thumbnail for the HAvatar technical paper

Citation

@article{zhao2023havatar,
  author = {Zhao, Xiaochen and Wang, Lizhen and Sun, Jingxiang and Zhang, Hongwen and Suo, Jinli and Liu, Yebin},
  title = {HAvatar: High-Fidelity Head Avatar via Facial Model Conditioned Neural Radiance Field},
  year = {2023},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  issn = {0730-0301},
  url = {https://doi.org/10.1145/3626316},
  doi = {10.1145/3626316},
  note = {Just Accepted},
  journal = {ACM Trans. Graph.},
  month = {oct},
  keywords = {parametric facial model, image-to-image translation, image synthesis, head avatar, neural radiance field}
}

Acknowledgements