Session Outline
This session focuses on the development of Voice2Face, EA’s state-of-the-art facial animation model. We introduce the model by discussing how we trained a generative model in a small data domain, while still reaching the quality expected from domain experts. Aiming to create a real-time application, we will then detail how we reduced the size of this model by several orders of magnitude without a large loss of quality.
Key Takeaways
- How to use generative modelling to overcome the one-to-many mapping in regression tasks
- How to leverage pre-trained models to train high quality generative models in small data domains
- How to employ knowledge distillation to reduce model size without a large loss of output quality
