RA-GAN: Fair and Race-Aware Face Aging

• RA-GAN is a race-aware face aging model built on a StyleGAN-V2 backbone with RACEpSp and a feature mixer to ensure identity, race, and target-age preservation.
• It addresses the generational age gap in kinship verification by transforming faces to a common age while maintaining accurate racial features and facial identity.
• Evaluated on metrics like racial accuracy, identity cosine similarity, and age mapping MAE, RA-GAN outperforms SAM-GAN and CUSP-GAN in fairness and verification tasks.

Searching arXiv for the specified paper to ground the response in the current record. RA-GAN is a face aging GAN model proposed for race-bias-free age transformation in kinship verification, where the central difficulty is the generational age gap between parent and child photographs and the absence of same-age images for many real pairs. The model is designed to produce aged faces that preserve identity, maintain racial characteristics, and match target ages, so that parent–child images can be transformed to a common age before verification. In the formulation reported for reliable kinship verification, RA-GAN is a StyleGAN-V2-based aging framework with two new modules, RACEpSp and a feature mixer, and it is evaluated primarily on racial accuracy, identity preservation, age mapping accuracy, and downstream kinship verification performance (Nazari et al., 18 Sep 2025).

1. Problem setting and fairness rationale

RA-GAN is motivated by two coupled issues in kinship verification: the age gap in parent–child photos and racial bias in face aging models (Nazari et al., 18 Sep 2025). Kinship verification takes two facial images and decides whether they are related, but in practical settings the images are captured at different times, so generational age differences introduce covariate shifts in wrinkles, skin texture, and bone structure. The same-age photos that would remove this shift are often unavailable.

The model is also motivated by the claim that existing face aging methods trained on racially imbalanced datasets often change racial characteristics during age transformation. The reported examples are FFHQ with approximately 70% White, LFW with approximately 77.5% White, and CelebA described as skewed to Asian. In the RA-GAN framing, such race drift harms likeness and identity preservation, which is consequential for kinship verification pipelines because they depend on comparing transformed images of the same subjects.

The fairness criterion used in the work is racial accuracy. It is defined as the classification accuracy of synthesized faces’ race labels, measured using a ResNet-34 race classifier trained on the balanced FairFace dataset. For each age group, the metric is the proportion of synthesized images whose predicted race matches the source image’s race label. Confusion matrices, precision, recall, and F1-score per race and age are also computed to assess fairness across classes.

The problem formulation is written with three simultaneous constraints on the synthesized output:

$\begin{align} x_{out} &= \text{RA-GAN} \left( x_{in} ,\ {age}_{t} \right) \notag \ s.t \quad & \text{identity}(x_{\text{out}) = \text{identity}(x_{\text{in}) \notag \ & \text{race}(x_{out}) = \text{race}(x_{in}) \notag \ & \text{age}(x_{out})= age_{t} \end{align}$

This formulation places race preservation alongside identity preservation and target-age consistency, rather than treating fairness as a post hoc diagnostic.

2. Architecture and latent-space design

RA-GAN is a conditional age transformation system built around a pre-trained, frozen StyleGAN-V2 generator $G$ that maps $W^+$ latent codes of shape $18 \times 512$ to RGB images (Nazari et al., 18 Sep 2025). The model is conditional on a target age $\alpha_t$, and age is injected as an additional channel:

$\alpha_t \sim \mathcal{U}(10,80)$

$x_{age}= x || \alpha_t$

The first trainable branch is the AgeEncoder $E_{age}$, which takes the 4-channel input and produces per-layer age style codes:

$$S_{age} = E_{age}(x_{age})\in \mathbb{R}^{(18,512)}$$

The second branch is RACEpSp $E_{race}$, a race-aware encoder that extends pSp by fusing race features with multi-scale facial features and then mapping them to $G$0:

$G$1

RACEpSp combines two frozen backbones. RaceNet is a ResNet-34 trained on FairFace, and PyramidNet is a pSp-style pyramid encoder providing multi-resolution facial feature maps. Feature fusion occurs through RaceMixer blocks. Each RaceMixer receives one pair of features from PyramidNet and RaceNet, uses transpose convolution for spatial alignment and a convolutional autoencoder for fusion, and includes a parallel scalar skip pathway to preserve critical facial information. The fused outputs $G$2 are produced at three scales, and each $G$3 is augmented by upsampled $G$4 to compensate information loss in pyramid convolutions before being processed by Map2Style heads.

The final trainable component is the feature mixer $G$5, which blends the age-conditioned and race-aware codes:

$G$6

The output image is then synthesized as

$G$7

and the complete mapping is

$G$8

The trainable parts are the RaceMixer blocks, Map2Style heads, AgeEncoder, and feature mixer. StyleGAN-V2, RaceNet, and PyramidNet are frozen. A common misconception is that RA-GAN follows standard adversarial training because of its name; the reported design explicitly states that no discriminator is trained and adversarial loss is not used. A plausible implication is that the method is closer to a regression-style latent editing framework built on a frozen generative prior than to a conventional end-to-end GAN training regime.

3. Objective functions and training procedure

The optimization objective combines reconstruction, identity preservation, age consistency, race preservation, and latent regularization terms (Nazari et al., 18 Sep 2025). The reconstruction loss is defined as

$G$9

the identity preservation loss uses the last CNN layer features of a pre-trained ArcFace ResNet-50,

$W^+$0

the age consistency loss uses VGG16 trained for age estimation,

$W^+$1

and the race preservation loss uses the last convolutional layer before the fully connected layer in the RaceNet model,

$W^+$2

The latent regularization term is reported in the paper as

$W^+$3

and the total objective is written with coefficients $W^+$4, $W^+$5, $W^+$6, $W^+$7, and $W^+$8.

Training uses a two-pass aging-and-reconstruction flow. For each batch $W^+$9, the forward path computes $18 \times 512$0 from $18 \times 512$1, $18 \times 512$2, $18 \times 512$3, and the feature mixer. A reconstructed phase then applies a second forward pass, $18 \times 512$4, to reconstruct $18 \times 512$5, computes a second objective $18 \times 512$6 toward $18 \times 512$7, and backpropagates both losses before stepping the optimizers. No discriminator update occurs; encoders and the mixer are optimized.

The reported training details are specific. The “first” optimizer uses learning rate $18 \times 512$8, betas $18 \times 512$9, and weight decay $\alpha_t$0. The reconstructed phase optimizer uses learning rate $\alpha_t$1, betas $\alpha_t$2, and weight decay $\alpha_t$3. Batch size is $\alpha_t$4, synthesis resolution is $\alpha_t$5, and the age range is 20 to 80 in steps of 10. Additional preprocessing includes GFPGAN denoising to $\alpha_t$6 to align latent distribution with StyleGAN-V2 and normalization of colored face images for stable training.

4. Data sources, evaluation protocol, and reported performance

The face aging model is trained on a race-balanced UTKFace-derived dataset across four races—Indian, White, Black, and Asian—with approximately balanced train and test counts (Nazari et al., 18 Sep 2025). The reported counts are: Train: Indian 1284, White 1235, Asian 1138, Black 1170; Test: Indian 316, White 295, Asian 274, Black 258. FairFace is used to train RaceNet. Kinship evaluation uses KinFaceW-I and KinFaceW-II with four relationships: Father-Son, Father-Daughter, Mother-Son, and Mother-Daughter.

For same-age transformation, each parent–child pair $\alpha_t$7 is first converted into full-face images through mirroring and an autoencoder: $\alpha_t$8, $\alpha_t$9. Both images are then transformed to a common target age $\alpha_t \sim \mathcal{U}(10,80)$0 using RA-GAN. Kinship verification is performed with D4ML (Distance and Direction-based Deep Discriminant Metric Learning) on the transformed pairs, using KinFaceW standard train/test splits and D4ML hyperparameters.

The reported fairness results center on racial accuracy per age group. RA-GAN outperforms SAM-GAN by 13.14 percentage points on average across all age groups from 20 to 80. At age 60, RA-GAN exceeds CUSP-GAN by 9.1 percentage points, while being comparable to CUSP-GAN at younger ages and better at older ages, with CUSP not evaluated beyond 60. Example racial accuracy entries are: age 20, RA-GAN 75.59, SAM-GAN 65.79, CUSP-GAN 78.21; age 60, RA-GAN 62.82, SAM-GAN 48.38, CUSP-GAN 53.71; age 70, RA-GAN 57.83, SAM-GAN 42.43. F1-score improvements over SAM-GAN averaged across ages are reported as Indian +23.28, Black +17.28, White +10.57, and Asian +7.42.

Identity preservation is measured by cosine similarity between ArcFace features of original and synthesized images. At age 60, the reported means and standard deviations are RA-GAN 0.413 (std 0.099), SAM-GAN 0.443 (0.096), and CUSP-GAN 0.353 (0.144). The paper states that RA-GAN preserves identity better than SAM-GAN and CUSP-GAN across age groups, while also describing the performance as nuanced: RA-GAN is comparable to SAM-GAN overall and better than CUSP-GAN in older ages.

Age mapping accuracy is measured by mean absolute error between target age and FACE++ estimated age on 1100 randomly sampled images per age group. RA-GAN consistently achieves lower MAE than SAM-GAN across all age groups. It is better than CUSP-GAN for ages greater than or equal to 30 and slightly worse at age 20. At age 60, the reported MAEs are RA-GAN 4.66 (std 3.44), SAM-GAN 6.31 (4.78), and CUSP-GAN 7.24 (5.38).

For downstream kinship verification, the reported gains after same-age transformation are relative to base full-face, non-transformed inputs. On KinFaceW-I, the gains are father-son +5.22 percentage points, father-daughter +5.12, mother-son +1.63, and mother-daughter +0.41. On KinFaceW-II, the gains are father-daughter +2.9, father-son +0.39, and mother-son +1.6.

5. Relation to SAM-GAN and CUSP-GAN

RA-GAN is positioned against SAM-GAN and CUSP-GAN in terms of both architectural emphasis and empirical behavior (Nazari et al., 18 Sep 2025). SAM-GAN adds age conditioning via pSp to edit StyleGAN latent codes but does not explicitly mitigate race bias; it is described as subject to a reconstruction–editing trade-off, with identity and race drift observed. CUSP-GAN emphasizes custom structure preservation and geometric consistency, but it lacks explicit race debiasing and is reported as limited to an age range less than or equal to 60.

The principal distinction claimed for RA-GAN is the explicit integration of race features and fairness losses. RACEpSp contributes race robustness by integrating RaceNet features, derived from a balanced race classifier, with PyramidNet features via RaceMixer. The feature mixer then balances age editing with identity and race preservation. The paper attributes improvements to the combined effect of RACEpSp, feature mixing, race loss, and balanced data.

Qualitative comparisons are consistent with this framing. Visual comparisons are reported to show that SAM-GAN often alters jaw shape, forehead proportions, and eye regions, with noticeable race drift at older ages. CUSP-GAN is described as maintaining structure at younger ages but tending to drastically change race in older ages such as 60. RA-GAN is reported to better model skin texture and aging effects, including wrinkles and sagging, while maintaining a coherent relationship between facial components.

An important methodological caveat is that the paper does not report explicit component-wise ablations. As a result, the individual contribution of RACEpSp, the feature mixer, the race preservation loss, and the balanced training set is not isolated experimentally. This limits causal attribution even though the combined system improves the reported fairness and downstream verification metrics.

6. Limitations, ethics, and practical use

The reported limitations are mainly about coverage, extreme-age behavior, and generalization (Nazari et al., 18 Sep 2025). Race categories are limited to four—Indian, White, Black, and Asian—and finer granularity such as ethnicity subgroups is not explored. Age distribution balancing within race reduces dataset size, and age ranges are discrete in 10-year steps. Identity similarity declines at ages 70 and 80, and pose variations remain challenging even with RaceNet features.

The paper also notes broader fairness and ethical considerations. While RA-GAN reduces race bias, fairness beyond the four races and across diverse capture conditions requires broader validation. Ethical use requires care when synthesizing faces of minors and managing privacy. These statements place the model in the category of bias-mitigation systems whose operating assumptions remain bounded by the composition of the training and evaluation data.

For practical usage, the released code is available at the GitHub repository linked from the paper. The reported dependencies include PyTorch, pre-trained StyleGAN-V2 weights, pSp encoder components, GFPGAN for denoising, a ResNet-34 race classifier trained on FairFace, ArcFace ResNet-50 for identity features, VGG16 DEX for age features, the FACE++ API for age estimation, and D4ML for kinship verification. The application procedure consists of preparing a race-balanced UTKFace-derived dataset and normalizing inputs; training or loading RaceNet and freezing RaceNet and PyramidNet; training the RA-GAN modules with the reported losses and coefficients; generating full-face images for KinFaceW-I or KinFaceW-II by mirroring cropped inputs and reconstructing them via the pSp plus StyleGAN-V2 autoencoder; selecting a target age $\alpha_t \sim \mathcal{U}(10,80)$1; running $\alpha_t \sim \mathcal{U}(10,80)$2 on parent and child images; and evaluating same-age pairs with D4ML, racial accuracy, ArcFace cosine similarity, and FACE++ age MAE as needed.

A plausible implication is that RA-GAN is best understood not only as a face aging model but also as a fairness-aware preprocessing stage for kinship verification, because its main empirical claims are tied to same-age transformation before D4ML rather than to unconstrained age synthesis alone.