Skip to main content
Back

Chinese Academy of Sciences' DiverAge Keeps Faces Consistent Across a Lifetime

DiverAge adds a training-free guidance module to diffusion-based face aging that keeps identity coherent across a full lifespan sequence while preserving visual diversity

  • DiverAge is a new face aging framework that generates diverse, lifespan-consistent aged faces using a training-free inference-time guidance module called CARR.
  • The core problem: pluralistic aging models generate diverse outputs per age group but let identity drift inconsistently across a full aging sequence spanning decades.
  • CARR (Cross-Age Identity Relation Regulator) jointly denoises all target age groups simultaneously, guided by a prior estimated from real same-identity cross-age face pairs.
  • No retraining required: CARR is a plug-in inference strategy with a single guidance strength hyperparameter, adding no trainable parameters to the base model.
  • Results show improved sequence-level ordinal reliability -- the identity similarity curve across age groups matches real aging patterns -- without sacrificing appearance diversity or image quality.
  • Applications include forensic age progression, cross-age biometric verification, and training data augmentation for face recognition systems; paper on arXiv, base model code on GitHub.

Face aging synthesis has a dirty secret: most models that generate diverse, realistic-looking aged faces will quietly break your identity if you ask them to age the same person across multiple decades at once. A new paper from researchers at the Chinese Academy of Sciences introduces DiverAge, a hierarchical framework built on top of diffusion autoencoders that tackles this problem head-on, without touching the training objective or adding any extra trainable parameters.

Why aging is harder than it looks

Face aging is not a simple image translation task. It is an ill-posed problem because multiple plausible aging patterns may correspond to a given input, since the human aging process is influenced by a variety of factors, including genetics and social environment. That means a good aging model should not produce a single deterministic output -- it should produce a distribution of plausible faces, all of which are consistent with how that person might realistically look at a given age.

The field has split into two camps trying to solve this. Most existing methods often produce one deterministic estimation, while newer work proposes pluralistic approaches to enhance the diversity of aging patterns. The problem is that pluralistic models introduce a new failure mode: when you generate a full lifespan sequence (say, ages 0-5, 10-20, 30-40, 50-60, 70+), each age group is sampled independently. The result is that the person's identity can drift inconsistently across the sequence -- the 30-year-old version might not look like a plausible older version of the 10-year-old version, even if both look individually realistic.

Comparison of deterministic aging, existing pluralistic aging, and DiverAge showing diversity and sequence reliability

This is the gap DiverAge is designed to close. The authors frame it as two separate requirements that existing work conflates:

  • Appearance-level diversity -- within a single age group, multiple plausible outputs should look different from each other (different wrinkle patterns, hair texture, etc.)
  • Sequence-level ordinal reliability