Title: DiverAge: Ensuring Reliable Pluralistic Face Aging via Cross-Age Identity Relation Guidance

Face aging is a critical component in forensic identity analysis, cross-age identity verification, and long-term biometric analysis. Because genetic, environmental, and lifestyle factors can lead to multiple plausible appearances for a single subject at a given target age, face aging is fundamentally a one-to-many generation task. However, relying solely on pluralism is inadequate for robust face aging. A effective model must balance appearance-level candidate diversity within individual age groups with sequence-level ordinal reliability across ordered age stages.

Current deterministic aging techniques can generate visually plausible age-progressed images but typically lack stochastic diversity. Conversely, while pluralistic aging methods introduce local appearance variations, they often fail to explicitly control the identity evolution across the entire aging sequence.

To address these challenges, we introduce DiverAge, a hierarchical pluralistic face aging framework grounded in diffusion autoencoding. DiverAge maintains appearance-level diversity by employing age-conditioned semantic modulation and stochastic diffusion decoding. To enhance sequence-level reliability, we propose the Cross-age Identity Relation Regulator (CARR). This is an inference-time guidance strategy that jointly denoises multiple target age groups. CARR leverages a Cross-age Identity Similarity (CIS) prior, derived from real same-identity cross-age pairs, to guide the process. It mitigates excessive cross-age identity drift through one-sided sampling-time guidance, achieving this without altering the training objective or adding extra trainable parameters.

Experimental results indicate that DiverAge successfully improves sequence-level ordinal reliability while preserving identity, age accuracy, image quality, and appearance-level diversity.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC