EGAvatar: Efficient GAN Inversion for Generalizable Head Avatar From Few-Shot Images
Summary
EGAvatar generates high-fidelity 3D head avatars from few images, overcoming limitations of current methods. This novel 3D Generative Adversarial Network (3DGAN) inversion framework ensures consistent identity and appearance for better avatar creation.
Area of Science:
- Computer Vision
- Computer Graphics
- Artificial Intelligence
Background:
- 3D generative models enable efficient avatar creation from images.
- Existing one-shot 3D head avatar reconstruction methods struggle with fidelity under limited input, causing distortions.
Purpose of the Study:
- To propose EGAvatar, an efficient 3D Generative Adversarial Network (3DGAN) inversion framework for high-fidelity head avatar reconstruction from few-shot images.
- To address shape distortions, expression deviations, and identity inconsistencies in current methods.
Main Methods:
- Developed a decoupling-by-inverting strategy using an animatable 3DGAN prior.
- Integrated coarse and offset 3D triplane representations within the 3DGAN.
- Implemented an expression-view disentanglement mechanism for consistent appearance across views and expressions.
Main Results:
- EGAvatar demonstrated superior qualitative and quantitative performance against state-of-the-art methods on multiple datasets.
- Achieved high-fidelity, generalizable 3D head avatars from significantly fewer input images.
- Showcased more efficient training and inference compared to existing approaches.
Conclusions:
- EGAvatar offers an effective solution for high-fidelity 3D head avatar reconstruction from limited input.
- The proposed framework enhances generalizability and consistency in avatar creation.
- EGAvatar represents a significant advancement in efficient and controllable 3D avatar generation.
