FaceArchive: Facial Image Tokenizer for Identity and Attribute Preserving Face Generation

Current approaches in face image generation typically focus on preserving identity or attributes exclusively. Identity preservation techniques maintain facial identity while enabling attribute manipulation, primarily serving personalized image generation applications. Conversely, attribute preservation methods retain facial characteristics while altering identity, commonly employed in face de-identification or dataset anonymization tasks. This dichotomy has created a significant gap in applications that require the simultaneous preservation of both aspects. In this paper, we present FaceArchive, a novel face tokenization framework that successfully disentangles and preserves both identity and attribute information from face images. Our architecture enables unprecedented control over face generation, allowing for either identity-preserving or attribute-preserving generation while maintaining high fidelity in both domains.