People do not age at one speed, and probably not at one speed per person either. A smoker's skin may run a decade ahead of her bone structure. Someone with a slow-changing face may still lose jawline definition on schedule. Current age estimators reduce all of this to one number. Some predict an age. Some predict an age with an uncertainty attached, as in label distribution learning, which outputs a probability spread over possible ages instead of a single value. Even models that split the network into several sub-networks, called experts, split them by chronological age bins. This proposal tests a different decomposition. It builds an age model whose experts are tied to distinct aging processes: facial geometry, periorbital texture, perioral and nasolabial texture, and forehead and skin tone. A small learned network, the router, decides how much each expert's reading counts at a given life stage. A person's aging rate then becomes a vector, with one rate per process, measured from that person's own photo history.
The core claim is falsifiable and has two parts. First, given two or more dated photos of a person, per-process rates estimated from them predict that person's later appearance and age better than a single scalar "ages fast or slow" offset. Second, these rates are stable within a person over time instead of being noise. If a scalar rate does as well, the multi-process view of non-uniform aging is wrong for faces at this data scale, and that is a useful negative result.
Why this, and not the obvious intersection
The obvious marriage of the two directions, a mixture of experts for age estimation, already exists. BridgeNet uses local regressors on overlapping age ranges, with a tree-shaped gating network whose lateral links keep neighbouring bins continuous (MORPH II MAE 2.38, FG-NET 2.56). Deep Regression Forests route features through soft decision trees to Gaussian leaves (MORPH 2.17, FG-NET 3.85). Both papers motivate this proposal: they show that life-stage non-stationarity matters. Both also partition capacity by age, and BridgeNet admits its gating targets are heuristic.
A second "obvious" gap is to route by inferred aging rate from a single image. It is ill-posed, because rate is a property of a trajectory, not of a snapshot. One image can show only how old someone looks relative to their birth date.
So the departure here is twofold. Experts are defined by which facial evidence they see, not by which ages they cover. Rate is estimated from longitudinal pairs of the same person, which MORPH II and FG-NET contain.
Adaptive Variance Distribution Learning (AVDL) learns a per-sample label spread through meta-learning. It finds that spread is widest in middle age (MORPH II 1.94, FG-NET 2.32 with ResNet-18). This is both motivation and the main threat: a monolithic model already absorbs much of the non-uniformity. We treat AVDL as the backbone to beat, not as a competitor to sidestep.
Model
A shared convolutional trunk, initialized as in AVDL from IMDB-WIKI pretraining, feeds K = 4 process experts. Each expert sees a restricted input. The geometry expert gets landmark coordinates and a low-pass whole-face crop. Each texture expert gets high-frequency patches from its own facial region.
The region restriction borrows the global-versus-local expert split of Face-MoGLE. That work separates holistic geometry from regional texture experts for face generation, and uses timestep-aware gating to shift weight between them. We use the same split for estimation, with a gate that depends on life stage instead of diffusion timestep.
Each expert outputs a scalar stage $s_k(x)$. The router outputs weights $g_k(x)$, and the predicted age is
$$\hat a(x) = \sum_{k=1}^{K} g_k(x)\, h_k\big(s_k(x)\big),$$
where $h_k$ is a learned monotone map from stage to years. Two losses train this. One is AVDL's distribution loss on $\hat a$. The other is a within-subject ranking loss that requires $s_k$ to increase between a person's earlier and later photos. That ranking loss is what makes each stage a progress coordinate, not just an arbitrary feature.
The router plays the role BridgeNet's chronological bins played. It should learn, unsupervised, that geometry dominates in childhood and texture in later adulthood. We will report this as a measured outcome, not an assumption.
We deliberately do not add the expert-decorrelation loss of MT-CRL, which decorrelates expert outputs to cut spurious sharing across tasks. Forcing decorrelation would manufacture the very independence of process rates we want to test. Process separation comes only from input restriction.
Personal rates
Take a subject with photos at ages $t_1 < t_2$. Their rate vector is $r_k = (s_k(x_2) - s_k(x_1))/(t_2 - t_1)$, normalized by the population-average slope at that age. A value of 1 means typical speed.
Given a new photo $x_3$, there are two uses. To forecast appearance, predict $s_k(x_3) \approx s_k(x_2) + r_k\,(t_3 - t_2)$. For personalized estimation, when $t_3$ is unknown, solve for the $t_3$ that best explains the observed $s(x_3)$ under the person's rates. Personalized estimation has chronological ground truth, so it can be scored with ordinary MAE.
The scalar-rate restriction fixes $r_k = r$ for all k. Its comparison against the full vector is the heart of the study.
Research Program
We will build the model and its estimation and forecasting procedures. We will train on subject-disjoint splits: MORPH II with its standard settings, and FG-NET with leave-one-person-out. CACD is used for cross-dataset transfer.
The analysis has four parts. We will report how router weights move with age. We will test whether rates estimated from photos 1→2 correlate with rates from photos 2→3 of the same person, which is test–retest reliability. We will measure whether process rates are correlated with each other across people; a single latent speed would show up as near-perfect correlation. Finally, we will check whether the vector adds predictive value beyond the scalar.
FG-NET's long spans test childhood and adulthood transitions. MORPH II has many subjects but short per-person spans, which makes rate estimates noisy. We will therefore report reliability as a function of time gap, not pool it away.
Baselines
The naive reference point uses no history at all: a population model that predicts age from the new photo alone. Here that model is our own architecture with rates fixed to 1. This tells us whether per-person information helps.
The strongest existing estimator is AVDL. It reports MAE 1.94 on MORPH II and 2.32 on FG-NET (AVDL). We will retrain it with the same ResNet-18 trunk, IMDB-WIKI pretraining, splits, and an equal hyperparameter-search budget. For the personalized task, we give it the same photo history through a scalar correction. We fit the person's offset, meaning apparent age minus chronological age, on earlier photos and carry it forward. We also fit a slope variant from two photos. This "AVDL plus scalar history" is the strongest competitor, because it uses everything our model uses except multi-process structure. Beating it means a statistically significant drop in personalized MAE on FG-NET and MORPH II (paired bootstrap over subjects), together with better forecasting of held-out stage vectors.
On single-photo MAE, we require only non-inferiority to AVDL's reported figures. Matching 1.94 and 2.32 shows the restricted-input design costs nothing. No reported result was found on Lacuna for personalized or history-conditioned age estimation on either benchmark, so for that task the only comparison is our reimplementation.
BridgeNet and DRF are run as alternative routing schemes inside the same trunk. They keep chronological-bin or soft-tree gating while receiving the same history correction. Their reported numbers (BridgeNet MORPH II 2.38, FG-NET 2.56; DRF MORPH 2.17, FG-NET 3.85) serve as sanity anchors.
Two ablations isolate the key ingredient. The first gives every expert the whole face, which removes the process definition while keeping the mixture. The second replaces the learned router with uniform weights, which tests whether life-stage gating matters.
Why not just…?
Why not just use AVDL with a per-person age offset? This is the real threat. AVDL's per-sample spread already captures much of the "this face is hard to place" effect. A scalar offset then captures "this person looks old for their age." If the stage vector forecasts no better than this, our hypothesis fails. The head-to-head personalized-MAE comparison and the cross-process correlation analysis settle it directly. We commit to reporting a null.
Why not just train region-specific regressors without a router? Region heads alone would give per-region ages. But the informativeness of geometry versus texture changes sharply between youth and senescence, as both BridgeNet and DRF argue. A fixed average would let an uninformative process add noise at the wrong stage. The uniform-router ablation tests exactly this. If it ties the learned router, we will drop the router and say so.
Why not just fine-tune a personal adapter, as MyTimeMachine does? MyTimeMachine learns a per-person latent offset from 10–50 photos for generation (Age MAE 7.7 against 25.9 for FADING plus DreamBooth). Our setting assumes two or three photos, which is too few for per-subject optimization and the regime where such fitting is fragile. An adapter also yields an opaque offset, not interpretable rates. We include a lightweight per-subject fine-tuning baseline on FG-NET, where photos per subject are most plentiful, at matched compute.
Proposed Experiments
The decisive check is personalized age estimation on FG-NET (leave-one-person-out) and MORPH II. The task is to predict the age of a held-out photo given two earlier dated photos. We compare the full rate vector, the scalar rate, the scalar offset applied to AVDL, and no history.
A second check is test–retest reliability of each $r_k$ across non-overlapping photo pairs. Rates that fail to correlate within a person are not a property of the person, whatever their predictive score.
A third check concerns transfer. We train on MORPH II and test forecasting on CACD. The question is whether the stage-versus-age curves and router schedule transfer across populations, given the cross-demographic disparities noted in the aging literature.
Finally, as a qualitative probe, we compare per-process rates on the small celebrity trajectory sets used by MyTimeMachine and The Aging Multiverse. Twelve subjects is far too few for statistics, so this is reported only as illustration. It shows whether the vector separates visibly different aging patterns. If it does, the vector could later serve as a conditioning signal for generators that currently branch on hand-specified lifestyle attributes.
Success means the vector beats the scalar on personalized MAE, rates show nonzero within-person reliability, and the router learns a geometry-to-texture shift with age without supervision. Failure on the first two means facial aging speed is, to measurable precision, one number. That would also justify the scalar-offset simplicity the field already uses.
References
- BridgeNet: A Continuity-Aware Probabilistic Network for Age Estimation.
- Deep Regression Forests for Age Estimation
- Adaptive Variance Based Label Distribution Learning for Facial Age Estimation
- Mixture of Global and Local Experts with Diffusion Transformer for Controllable Face Generation
- Improving Multi-Task Generalization via Regularizing Spurious Correlation
- MyTimeMachine: Personalized Facial Age Transformation
- The Aging Multiverse: Generating Condition-Aware Facial Aging Tree via Training-Free Diffusion