🤖 AI

DiverAge: Reliable Pluralistic Face Aging with Cross-Age Identity Relation Guidance

The paper proposes DiverAge, a hierarchical pluralistic face aging framework based on diffusion autoencoding that utilizes a novel Cross-age Identity Relation Regulator (CARR) to simultaneously ensure appearance-level diversity and sequence-level ordinal reliability without requiring additional training parameters.

Original authors: Yueying Zou, Peipei Li, Qianrui Teng, Dianyan Xu, Zekun Li

Published 2026-06-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yueying Zou, Peipei Li, Qianrui Teng, Dianyan Xu, Zekun Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict what a person will look like when they are older. In the past, computer programs tried to do this by drawing a single, straight line from "now" to "then." They would say, "If you are 20 today, here is exactly what you will look like at 40." But the problem is, life isn't that predictable. The same person could age in many different ways depending on their genes, lifestyle, and environment. One 40-year-old might have deep wrinkles, while another might look smooth.

This paper introduces a new system called DiverAge (Diverse Aging) that solves two big problems with current technology:

1. The Problem of "One-Size-Fits-All" vs. "Chaos"

The Old Way (Deterministic): Think of this like a photocopier that only makes one copy. If you ask it to show a person at age 40, it gives you the exact same face every time. It misses the natural variety of how people actually age.
The "Chaos" Way (Existing Pluralistic Methods): Other newer methods try to fix this by making many different versions (pluralism). However, they act like a chaotic artist. They might make four different versions of a 40-year-old, but if you look at the whole timeline (age 20, 30, 40, 50), the faces might jump around weirdly. The person might look like a different family member at age 40 than they did at age 30. The "story" of their aging doesn't make sense.

2. The DiverAge Solution: A "Smart Storyteller"

DiverAge is like a smart storyteller who understands two rules:

Variety is good: At any single age (like 40), the person should have different possible looks (wrinkles here, smooth skin there).
The story must make sense: If you line up all the ages from young to old, the faces must look like the same person evolving naturally. They shouldn't suddenly look like a stranger in the middle of the timeline.

How It Works (The Metaphors)

The "Diffusion Autoencoder" (The Artist's Canvas)
Think of the core technology as a painter who starts with a blurry sketch and slowly refines it into a clear picture. This painter is naturally "stochastic," meaning if you ask them to paint the same scene twice, they will add slightly different details (like different wrinkles or skin texture) each time. This handles the variety part.

The "CARR" (The Editor with a Rulebook)
This is the paper's big innovation. Imagine you have a rulebook based on real-life data (like a massive photo album of real families growing up). This rulebook tells the computer: "Usually, a person's face at age 30 is very similar to age 35, but less similar to age 60."

The system uses a tool called CARR (Cross-age Identity Relation Regulator). It acts like a strict editor during the painting process:

It looks at all the ages being generated at once.
It checks the "rulebook" (called the CIS Prior) to see if the faces are drifting too far apart.
If the computer tries to make the 40-year-old look too different from the 30-year-old (breaking the family resemblance), CARR gently nudges the image back to stay on track.

Crucially, this editor doesn't need to retrain the painter. It just gives advice while the picture is being made.

What They Found

The researchers tested DiverAge and found:

It keeps the variety: You can still get different-looking faces for the same age (different wrinkles, skin textures).
It fixes the timeline: The sequence of faces from young to old now flows naturally. The "identity" stays consistent, and the jump between ages feels realistic, not abrupt.
It doesn't break the basics: The faces still look real, the age is accurate, and the person still looks like themselves.

In short, DiverAge teaches computers that aging isn't just about changing a face; it's about telling a reliable story where the character changes in many plausible ways, but remains the same person throughout the journey.

Technical Summary: DiverAge

Problem Statement

Face aging is critical for long-term biometric analysis, cross-age identity verification, and forensic applications. However, it is inherently a one-to-many generation problem: due to genetic, environmental, and lifestyle factors, a single subject can exhibit multiple plausible appearances at a target age.

Existing methods fail to address the full scope of "reliable" pluralistic aging:

Deterministic methods (e.g., FADING, FaceTT) produce visually plausible, identity-preserving faces but lack stochastic diversity, often generating repeated outputs for the same input and target age.
Pluralistic methods (e.g., PADA, ADFD) introduce local appearance variations (wrinkles, texture) but often fail to regulate the sequence-level ordinal reliability. They generate diverse samples for individual ages but do not ensure that the full aging sequence follows the natural, monotonic identity-similarity patterns observed in real cross-age data.
Ordered trajectory methods (e.g., Lifespan) model aging along an axis but typically map identities to single deterministic trajectories, ignoring the biological reality of multiple plausible aging paths.

Conventional evaluation metrics (image quality, aging accuracy, identity preservation) are frame-level and fail to measure whether a generated lifespan sequence maintains statistically reliable cross-age identity relationships.

Methodology: DiverAge

The authors propose DiverAge, a hierarchical pluralistic face aging framework based on diffusion autoencoding. It addresses reliability at two levels: appearance-level candidate diversity and sequence-level ordinal reliability.

1. Hierarchical Framework

Appearance-Level Diversity: DiverAge inherits a stochastic diffusion autoencoding backbone. By using stochastic diffusion decoding and age-conditioned semantic modulation, it generates multiple plausible candidates (varying in wrinkles, skin texture, and pigmentation) for the same input and target age.
Sequence-Level Reliability: To ensure the generated lifespan sequence follows natural aging patterns, the framework introduces a Cross-age Identity Relation Regulator (CARR).

2. Cross-age Identity Similarity (CIS) Prior

The authors construct a statistical prior from real same-identity cross-age pairs (using datasets like FFHQ-AT).

They define Cross-age Identity Similarity (CIS) as the cosine similarity between face recognition embeddings of the same person at different ages.
This prior ( $Z_{CIS}$ ) characterizes how identity similarity naturally decays as the age gap increases. It serves as a population-level reference for expected similarity between ordered age groups.

3. Cross-age Identity Relation Regulator (CARR)

CARR is an inference-time guidance strategy that does not modify the training objective or introduce extra trainable parameters.

Mechanism: During DDIM sampling, CARR jointly denoises multiple target age groups. It computes the generated cross-age identity similarity matrix and compares it against the $Z_{CIS}$ prior.
Loss Function: It employs a one-sided hinge guidance loss ( $L_{CARR}$ ). This loss penalizes age pairs where the generated identity similarity falls below the expected real-data prior (excessive identity drift), while leaving pairs with sufficient similarity unchanged.
Update: The latent variables are updated via gradient descent using this loss, guiding the sampling process toward ordinally coherent sequences without altering the underlying diffusion model.

4. Evaluation Metrics

The paper introduces Age Progression Rationality (APR) metrics to evaluate sequence-level reliability:

CIS Deviation ( $E_{CIS}$ ): Measures the deviation of the generated similarity matrix from the real-data prior.
Identity Drift Error ( $E_{drift}$ ): A one-sided error penalizing generated pairs with similarity lower than the prior.

Key Contributions

Problem Formulation: The paper formulates reliable pluralistic face aging as a hierarchical problem requiring both appearance-level diversity and sequence-level ordinal reliability, moving beyond conventional frame-level evaluation.
CIS Prior & Metrics: It constructs a CIS prior from real data to statistically characterize cross-age identity evolution and introduces APR-related metrics (CIS deviation, drift error) for sequence-level evaluation.
DiverAge Framework: It proposes a diffusion-based framework equipped with CARR, which injects the CIS prior into DDIM sampling to improve sequence reliability without retraining or adding parameters.
Empirical Validation: Extensive experiments demonstrate that DiverAge improves sequence-level ordinal reliability while maintaining identity preservation, age accuracy, image quality, and appearance-level diversity.

Experimental Results

Experiments were conducted on the FFHQ-AT dataset, comparing DiverAge against deterministic baselines (LATS, SAM, CUSP) and pluralistic methods (PADA).

Frame-Level Quality: DiverAge achieves competitive performance in Age MAE, Identity Similarity, and FID, comparable to state-of-the-art deterministic methods when evaluated in a single-output setting.
Sequence-Level Reliability: DiverAge significantly reduces the APR Error (0.0562) compared to the baseline PADA (0.0857) and other methods, indicating that generated sequences better follow the real-world identity-similarity decay pattern.
Diversity: The method maintains high LPIPS-div scores (0.1091), confirming that the introduction of sequence-level guidance does not suppress the stochastic appearance variations required for pluralistic aging.
Ablation Studies: Results confirm that the age-conditioned latent prior improves age controllability, while CARR specifically corrects excessive identity drift, leading to the best overall trade-off.

Significance and Claims

The paper claims that DiverAge represents a shift from treating face aging as a single-output editing task or a purely stochastic generation task to a reliable trajectory generation problem.

Reliability: The core significance lies in ensuring that pluralistic aging outputs are not just diverse but statistically consistent with real-world cross-age identity evolution.
Efficiency: By using inference-time guidance (CARR) rather than retraining, the method achieves sequence-level reliability without the computational cost of training new models or adding parameters.
Application: The framework is positioned as a robust solution for long-term biometric analysis and forensic identity analysis, where the consistency of an aging sequence is as critical as the quality of individual frames.

The authors acknowledge limitations, noting that the CIS prior relies on face recognition embeddings (potentially inheriting biases), that CARR requires extra gradient computation during inference, and that the method currently focuses on facial attributes rather than contextual factors like clothing or hairstyle.