Identity-Preserving Aging and De-Aging of Faces in the StyleGAN Latent Space

Technical Summary: Identity-Preserving Aging and De-Aging of Faces in the StyleGAN Latent Space

Problem Statement
Face aging and de-aging using generative AI is critical for applications in forensics, security, and media. However, state-of-the-art methods typically rely on Conditional GANs, Diffusion models, or Visual Language Models that require complex training procedures, extensive labeled datasets, and specific conditioning (e.g., target age labels or text prompts). These approaches often struggle with controllability and, crucially, fail to guarantee identity preservation. Existing identity preservation mechanisms frequently depend on a single face recognition model embedded in the loss function, which introduces bias and limits the robustness of the output across different recognition systems. Furthermore, navigating the continuous latent space of GANs to find a consistent "age direction" without degrading identity remains a significant challenge.

Methodology
The authors propose a data-efficient strategy that operates directly within the latent space of a pre-trained StyleGAN2 (specifically the $W$ space), avoiding the need to re-train generators or encoders. The methodology consists of three core components:

1. Linear Latent Editing via Support Vector Regression (SVR):
   Instead of training new networks, the authors project real face images into the StyleGAN2 $W$ latent space. They employ a Linear Support Vector Regressor (SVR) to fit a hyperplane that maps latent vectors ($\tilde{w}$) to their corresponding age labels ($y$). By projecting the SVR parameters onto a latent vector, they derive a normalized unit vector ($\hat{\lambda}$) representing the global "age synthesis direction." Aging or de-aging is achieved by moving along this direction using a scalar step $s$:
   $$w_1 = w_0 + s\hat{\lambda}$$
   To map the scalar $s$ to specific target ages (years), the authors fit polynomial curves to the relationship between the scalar step and the estimated apparent age, allowing for the calculation of precise scalar offsets ($\Delta s$) for any desired age transformation.

2. Feature Selection for Identity Preservation:
   The linear baseline approach modifies all latent components indiscriminately, which can degrade identity. To mitigate this, the authors introduce a weighting vector $\Phi$ that selectively scales the age direction components:
   $$w_1 = w_0 + \Phi \odot s\hat{\lambda}$$
   The vector $\Phi$ is computed using feature selection strategies based on Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA).

   • PCA: Identifies components explaining 95% of the variance in identity-labeled latent vectors.
   • LDA: Computes discriminative components for both identity and age groups. The authors reconstruct projected latents back to the original space and use distance metrics (Mean Squared Error, 1-D Wasserstein distance, and Covariance) to determine which components are critical for identity versus age.
   • The final weighting vector combines components relevant to age only, identity only, and those shared by both, controlled by hyperparameters $\alpha$ and $\beta$.

3. Empirical Validation and Dataset Generation:
   The method is evaluated using two state-of-the-art face recognition systems (IResNet50 and EdgeFace-S) and an age estimator (MobileNetV2). The authors verify identity preservation by measuring the cosine distance between synthesized and original faces across different age groups. They also generate a fully synthetic dataset using Langevin sampling to create 20,000 unique identities, each with 10 age variations, to serve as a benchmark for cross-age recognition and synthetic image detection.

Key Contributions

1. Streamlined Latent Editing: A baseline approach for face age synthesis using linear editing within StyleGAN2's $W$ space, eliminating the need for conditioning networks and large labeled datasets.
2. Feature Selection Strategies: The proposal and evaluation of PCA and LDA-based feature selection mechanisms that effectively improve identity preservation while synthesizing apparent age changes.
3. Identity Preservation Analysis: A set of experiments and visualization techniques that identify thresholds where identity begins to degrade during age modification, providing empirical bounds for safe editing.
4. Practical Toolset: A release of linear age directions, component weights, fitted curves per age group, and code to calculate the necessary scalar steps for specific target ages.
5. Synthetic Dataset: The generation and release of a public dataset containing 20,000 synthetic identities with 10 age variations each, addressing privacy concerns associated with real cross-age data.

Results

• Age Estimation: The relationship between the scalar step $s$ and apparent age is non-linear, particularly for subjects in extreme age groups (very young or very old). Polynomial fitting successfully models this relationship, enabling precise age control.
• Identity Preservation Bounds: The linear baseline allows for significant age gain (aging children or de-aging seniors) but shows asymmetry; it is less effective at de-aging children or aging seniors. The choice of face recognition system significantly impacts the observed limits (e.g., EdgeFace-S demonstrates a wider valid age range than IResNet50).
• Feature Selection Efficacy: All feature selection methods outperform the linear baseline. Specifically, LDA using Covariance and PCA showed the most improvement. LDA with Covariance performed slightly better for children, while PCA was more robust for seniors. These methods allow for larger age transformations while maintaining a 75% face verification rate.
• Comparison: The proposed method qualitatively and quantitatively outperforms the SAM (Style-based Age Manipulation) approach, particularly across larger age gaps.

Significance
The paper demonstrates that identity-preserving age edits in StyleGAN2 can be achieved without re-training generators, relying instead on a simple SVR-based latent direction and strategic feature selection. By establishing practical scalar-to-age mappings and identifying the limits of identity preservation, the work provides a more controllable and data-efficient alternative to complex conditional models. The release of a fully synthetic dataset and toolset aims to facilitate benchmarking for cross-age face recognition, age assurance systems, and synthetic image detection, while addressing the privacy and bias issues inherent in using real-world cross-age data.