StyleID: A Perception-Aware Dataset and Metric for Stylization-Agnostic Facial Identity Recognition

Researchers introduced StyleID and StyleBench-S, a perception-aware benchmark for facial identity recognition under stylization.

The study finds that standard identity encoders often degrade on heavily stylized faces, missing cues humans still use to judge identity.

Using human 2AFC judgments and psychometric curves, the team calibrates encoders so their scores better match human perception across styles.

The result is a more reliable evaluation setup for stylization-agnostic identity recognition, addressing a major failure mode in face recognition.