EXPL-FR Explains Face Recognition via Vision-Language Alignment
August 20, 2026
EXPL-FR uses a lightweight adapter to align a vision-language model's image encoder with a frozen face recognition embedding space. This allows users to query similarity scores using text prompts for 22 different semantic categories without retraining the core FR model.
HOW THIS AFFECTS YOU
●
builderYou can provide semantic interpretability for black-box face recognition systems.
●
researcherThis demonstrates effective cross-modal alignment between VLMs and specialized embedding spaces.