Source-linked AI summary
Face Aging With Conditional Generative Adversarial Networks
Grigory Antipov, Moez Baccouche, Jean-Luc Dugelay
TL;DR
The paper addresses identity loss in GAN-based face alteration while seeking natural, age-conditioned face synthesis. It proposes Age-cGAN with identity-preserving latent-vector optimization, and reports objective evaluations demonstrating the method’s potential. The approach reconstructs an input face before changing the generator’s age condition, while its reconstruction remains an approximation and further improvement is planned.
Problem
GAN-based face modification can lose the original person’s identity, although face aging requires changing age while keeping identity intact.
Method
Age-cGAN reconstructs an input face by optimizing its latent vector with identity preservation, then performs aging by changing the generator’s age condition.
Results
The method’s identity-preserving reconstruction outperforms pixelwise optimization subjectively and objectively, while Age-cGAN disentangles identity-related information from age conditions.
Takeaways & Limitations
The identity-preserving optimization is presented as universal for face aging and other alterations such as adding beards or sunglasses.
Takeaways & Limitations
Age-cGAN approximates rather than exactly reconstructs all real-life face images, and combining pixelwise with identity-preserving optimization remains future work.
Abstract
from arXiv · showhide
It has been recently shown that Generative Adversarial Networks (GANs) can produce synthetic images of exceptional visual fidelity. In this work, we propose the GAN-based method for automatic face aging. Contrary to previous works employing GANs for altering of facial attributes, we make a particular emphasize on preserving the original person's identity in the aged version of his/her face. To this end, we introduce a novel approach for "Identity-Preserving" optimization of GAN's latent vectors. The objective evaluation of the resulting aged and rejuvenated face images by the state-of-the-art face recognition and age estimation solutions demonstrate the high potential of the proposed method.
1. INTRODUCTION
Traditional face-aging methods either generalize across age groups and lose personalized information or require costly age sequences. The paper therefore introduces Age-cGAN and identity-preserving latent optimization for high-quality, age-conditioned face synthesis.
- Limitations of traditional methods: Traditional prototyping methods are simple and fast but discard personalized information, producing unrealistic images.They transform inputs using differences between average faces in predefined age groups.
- Limitations of traditional methods: Modeling approaches simulate individual aging mechanisms but often require costly, wide-range age sequences of the same person.These models target muscles, skin, and skull aging for a particular individual.
- Motivation for the proposed method: Previous GAN-based face-modification methods can produce realistic alterations but often lose the original person’s identity.The paper motivates identity preservation as the central challenge for GAN-based aging and related face modifications.
- Contributions: Age-cGAN is designed to generate high-quality synthetic face images within required age categories.The contribution targets age-conditioned generation rather than only binary older-or-younger transformations.
- Contributions: A novel latent-vector optimization approach reconstructs an input face while preserving the original person’s identity.This addresses the identity loss identified in earlier GAN-based face-modification methods.
2. PROPOSED METHOD
The method uses Age-cGAN to reconstruct an input face, then changes its age condition while preserving identity through latent-vector optimization.
- Face aging pipeline: The method first finds a latent vector z* that reconstructs the input image as G(z*, y0), then switches the condition to generate the target-age image.This separates identity reconstruction from age alteration in two pipeline steps.
- Age-cGAN: Age-cGAN generates face images conditioned on six age categories and is used as the generative model for face aging.The categories are 0–18, 19–29, 30–39, 40–49, 50–59, and 60+ years.
- Initial latent approximation: Because cGANs lack an explicit inverse mapping from images to latent vectors, an encoder E provides an initial latent approximation for reconstruction.E is trained on 100K synthetic image–latent-vector pairs using Euclidean distance between estimated and ground-truth latent vectors.
- Identity preservation: The initial encoder-based approximations can produce plausible reconstructions but may lose the original identity in about 50% of cases.GANs also cannot exactly reproduce every detail of real face images, so the input is generally approximated rather than exactly reconstructed.
- Identity-Preserving optimization: The proposed Identity-Preserving optimization minimizes the Euclidean distance between face-recognition embeddings of the original and reconstructed images.Unlike pixelwise optimization, the objective focuses on identity-related information rather than unnecessary facial details that can increase blur.
- Identity-Preserving optimization: Differentiable generator and recognition networks allow L-BFGS-B optimization, initialized from the encoder’s latent approximation; z*IP is reported to preserve identity better than z*pixel.The recognition network is an internal implementation of FaceNet.
3. EXPERIMENTS
Experiments evaluate Age-cGAN’s age-conditioned generation and identity-preserving reconstruction on the IMDB-Wiki cleaned dataset. The comparisons show that identity-preserving optimization improves reconstruction identity while changing age through the generator’s condition.
- 3.1. Dataset: Age-cGAN was trained on about 120K IMDB-Wiki cleaned images, using 110K for training and 10K for identity-preserving reconstruction evaluation.
- 3.2. Age-Conditioned Face Generation: Each latent vector generates faces across six age conditions, with latent vectors encoding identity-related attributes and y encoding age.
- 3.3. Identity-Preserving Face Reconstruction and Aging: Optimized reconstructions are closer to original images than initial reconstructions, while Pixelwise optimization better preserves superficial details and Identity-Preserving optimization better represents identity traits.
- 3.3. Identity-Preserving Face Reconstruction and Aging: 82.9% of OpenFace outputs recognized the original person after Identity-Preserving optimization, compared with about half for initial reconstructions and marginal improvement for Pixelwise optimization.
- 3.3. Identity-Preserving Face Reconstruction and Aging: After identity-preserving reconstruction, replacing the original age condition y0 with ytarget produces the target-age face, supporting both aging and rejuvenation examples.
4. CONCLUSIONS AND FUTURE WORK
The paper concludes that Age-cGAN enables synthetic face aging through identity-preserving latent optimization and condition changes. It identifies dataset augmentation, cross-age recognition robustness, and combined optimization as future directions.
- The proposed method reconstructs an input face by optimizing a latent approximation, then performs aging by changing the generator’s age condition.
- Identity-Preserving latent optimization preserves the original person’s identity and is intended to extend to alterations such as beards and sunglasses.
- Synthetic aging is proposed for face-dataset augmentation and improving face-recognition robustness in cross-age scenarios, as future work.
- The authors plan to improve reconstruction by combining Pixelwise and Identity-Preserving approaches in one optimization objective.