NAMESAKES: Probing Identity Memorization in Text-to-Image Models

The article introduces the NAMESAKES dataset, comprising over one thousand names and faces of public figures, to address identity memorization in text-to-image (T2I) models. A novel black-box behavioral probe is presented that can distinguish between memorized and fabricated identities without requiring access to training data or ground-truth images. This development is significant for practitioners as it enhances the understanding of privacy implications in T2I models and offers a method to evaluate identity memorization across different model architectures.