Research
Independent research · Jun 2026 – Aug 2026
Do Face Embeddings Predict Attractiveness?
- Compared 11 face-embedding families on ~7,800 rated faces
- Identity-trained embeddings predicted attractiveness worse
- CLIP best (r = 0.94), Fisherface worst (r = 0.32)
About
Benchmarked 11 face-embedding families (deep identity CNNs, classical methods, CLIP, DINOv2) against 3 human attractiveness-rating datasets totaling ~7,800 rated faces, using frozen embeddings and ridge regression. Found that embeddings trained to capture identity did consistently worse at predicting attractiveness than those that were not, pointing to a tradeoff between pose invariance and attractiveness predictivity. CLIP led every dataset (r = 0.94 on SCUT-FBP5500), while Fisherface—the most aggressively identity-separating method tested—finished last at a mean r = 0.32.
- Python
- PyTorch
- scikit-learn
- Computer Vision