Overview
Implemented a Face-to-Character PoC that estimates character parameters from facial images. Rather than connecting the paper's non-differentiable game renderer directly, I used a GAN as a differentiable proxy to experiment with parameter optimization.
Conversion Examples
Original
Game Character
Animation
Key Achievements
- Prepared data, implemented and trained the GAN architecture, and performed qualitative result evaluation
- Compared GAN, VAE, and autoencoder approaches; selected GAN for the strongest visual quality at the time
- Game-engine integration was not completed; the work remains a PoC
Tech Stack
- Framework: PyTorch
- Vision: Face Detection, Landmark Extraction
- Engine: Unreal Engine