How It Works
Dynaface builds on three neural networks, all run locally through ONNX Runtime:
| Model | Role | License |
|---|---|---|
| BlazeFace (short-range) | Face bounding-box detection | Apache-2.0 (MediaPipe) |
| SPIGA (WFLW-98) | 98-point landmarks and head pose | BSD-3-Clause |
| U²-Net | Background/saliency removal for lateral views | Apache-2.0 |
Research Using Dynaface
Dynaface has been used in peer-reviewed medical research to quantify oral-ocular synkinesis and to study how AI-derived facial metrics relate to patient-reported outcomes. If you use Dynaface in academic work, please cite:
- Renne, A., Heaton, J., & Boahene, K. D. O. (2026). Associations of AI-based facial metrics with patient-reported outcomes in idiopathic facial paralysis. Laryngoscope. Advance online publication. https://doi.org/10.1002/lary.70417
- Renne, A., Heaton, J., Derakhshan, A., Nellis, J. C., Desai, S. C., & Boahene, K. D. (2025). Use of dynamic, automated facial analysis in quantifying oral-ocular synkinesis. Facial Plastic Surgery & Aesthetic Medicine. https://doi.org/10.1177/26893614251395737
- Berges, A. J., Renne, A., Heaton, J., Leung, D. G., & Boahene, K. D. (2025). Facial weakness in facioscapulohumeral muscular dystrophy: Objective and patient-reported measures to guide reconstructive interventions. Facial Plastic Surgery & Aesthetic Medicine, Article 26893614251407675. https://doi.org/10.1177/26893614251407675
A BibTeX entry is available in CITATION.bib. Please also cite the original authors of the three neural networks listed above; their references are in the project README.