What is it about?
A model compression framework in the area of 3D facial animation. We compressed a large 3D facial model so that it can run on a user device in real time, while still keeping high performance. The large model has about 1000 million parameters and consumes ~5GB of GPU memory to generate animation, while our compressed on-device model has ~0.3 million parameters and takes 3.4MB of CPU memory. The real-time model has a latency of ~81ms.
Featured Image
Why is it important?
The state-of-the-art 3D facial animation models of recent years are extremely large due to the use of pre-trained speech encoders, making them unable to run on a user device or in real time. We are interested in next-gen game applications like real-time 3D avatars or animating user-specific dialogues on the user's device to bring more personalized, unique, and immersive game experiences to players. This work lays the foundation for realizing our long-term vision.
Perspectives
We hope this research challenges the current trend of building ever-larger models, and instead encourages more efficient, lightweight alternatives. As researchers in a game company, our goal is to apply AI in practical, product-driven ways.
Zhen Han
Electronic Arts Inc
Read the Original
This page is a summary of: Tiny is not small enough: High quality, low-resource facial animation models through hybrid knowledge distillation, ACM Transactions on Graphics, July 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3730929.
You can read the full text:
Resources
Contributors
The following have contributed to this page
