Identity-Preserving Aging and De-Aging of Faces in the StyleGAN Latent Space
Date:
Two-minute poster spotlight in the main conference, followed by the poster session.
Most face aging methods rely on heavy conditional GANs, diffusion models, or vision-language models, and do not guarantee that the identity is preserved. We find an “age direction” in the StyleGAN2 latent space with a linear SVR, constrain the edit to an identity-preserving subspace through PCA/LDA-based feature selection, and map latent offsets to specific target ages. With this method we released a fully synthetic dataset of 20k identities, each at 11 ages, for benchmarking cross-age face recognition, age assurance, and synthetic image detection.
Authors: Luis S. Luevano, Pavel Korshunov, Sébastien Marcel.