Source-linked AI summary
Age Progression/Regression by Conditional Adversarial Autoencoder
Zhifei Zhang, Yang Song, Hairong Qi
TL;DR
Existing face-aging methods often require paired faces across ages and a query image with a known age, limiting use with unpaired and unlabeled data. The paper introduces CAAE, which encodes personality and conditions generation on age while traversing a learned face manifold. The authors report flexible, effective, robust, and more photo-realistic age progression and regression, including preservation of personality across age changes.
Problem
Existing methods commonly require paired samples across ages and labeled query faces, while age regression and realistic rejuvenation remain challenging.
Method
CAAE encodes a face into a personality-preserving latent vector, concatenates an age label, and uses a generator plus two discriminators to map it onto a traversable face manifold.
Results
The authors report photo-realistic age progression and regression from arbitrary query faces without true-age labels, with personality preservation and robustness to pose, expression, and occlusion.
Takeaways & Limitations
CAAE provides a flexible manifold-based framework for manipulating apparent age in both directions while preserving facial identity-related personality features.
Takeaways & Limitations
The framework relies on the assumption that face images lie on a high-dimensional manifold whose traversable directions separate age and personality.
Abstract
from arXiv · showhide
"If I provide you a face image of mine (without telling you the actual age when I took the picture) and a large amount of face images that I crawled (containing labeled faces of different ages but not necessarily paired), can you show me what I would look like when I am 80 or what I was like when I was 5?" The answer is probably a "No." Most existing face aging works attempt to learn the transformation between age groups and thus would require the paired samples as well as the labeled query image. In this paper, we look at the problem from a generative modeling perspective such that no paired samples is required. In addition, given an unlabeled image, the generative model can directly produce the image with desired age attribute. We propose a conditional adversarial autoencoder (CAAE) that learns a face manifold, traversing on which smooth age progression and regression can be realized simultaneously. In CAAE, the face is first mapped to a latent vector through a convolutional encoder, and then the vector is projected to the face manifold conditional on age through a deconvolutional generator. The latent vector preserves personalized face features (i.e., personality) and the age condition controls progression vs. regression. Two adversarial networks are imposed on the encoder and generator, respectively, forcing to generate more photo-realistic faces. Experimental results demonstrate the appealing performance and flexibility of the proposed framework by comparing with the state-of-the-art and ground truth.
1. Introduction
Face age progression and regression aims to change apparent age while preserving personal features, but existing methods often require difficult paired data and labeled query ages. The paper proposes CAAE, a generative manifold-based framework intended to perform both directions without those requirements.
- Face age progression predicts future looks, while regression estimates previous looks, with both aiming to preserve personalized facial features.
- Existing methods commonly require paired images of the same person at different ages and a query image labeled with its true age.The Morph dataset captured an average of only 164 days between images for each individual.
- Age regression is especially difficult because many methods remove surface texture and therefore struggle to produce realistic baby-face predictions.
- CAAE treats faces as lying on a high-dimensional manifold and traverses an age direction to generate different ages while preserving personality.
- The proposed framework performs age progression and regression, avoids paired training samples and labeled test faces, and uses disentangled age and personality representations.The paper also claims photo-realistic outputs and robustness to pose, expression, and occlusion.
2. Related Work
Related work includes physical and prototype-based aging methods, while GANs provide a generative alternative for realistic image synthesis. Existing approaches face data, modeling, or realism limitations, particularly for rejuvenation.
- Physical model-based methods represent biological aging through facial mechanisms such as muscles, wrinkles, and structure.They require large datasets covering long age spans for each individual, which are difficult to collect.
- Prototype-based methods group faces by age and use differences between average prototypes as aging patterns.Averaging can lose personalized details such as wrinkles.
- Face age regression remains challenging because surface-based methods mainly remove learned textures and cannot produce photo-realistic baby-face predictions.
- GANs generate samples through a generator and discriminator trained in competition, but original GAN training can be unstable and produce noisy images.
3. Traversing on the Manifold
The framework assumes that age and personality correspond to separable directions on a face manifold. It encodes personality into a latent vector, combines it with age labels, and traverses the age direction before generating faces.
- The method maps faces to a lower-dimensional latent space because directly modeling and traversing the high-dimensional face manifold is difficult.
- An encoder E maps input faces x1 and x2 to latent vectors z1 and z2 representing personality features.
- Age labels are concatenated with personality vectors to form latent points [z1, l1] and [z2, l2].
- Traversing along the age axis changes age while preserving personality, whereas traversing across both axes changes age and personality.
- A generator G maps traversed latent points back to the face manifold, producing face images for age progression and regression.
4. Approach
The CAAE learns a manipulable face manifold by encoding personality separately from age, then uses adversarial regularization to produce realistic age-conditioned faces. Its two discriminators smooth latent-space traversal and improve output realism, supporting both age progression and regression.
- Network architecture: CAAE encodes an input face into personality vector z and conditions generator G on age label l to produce an age-specific face.The encoder avoids random sampling so generated faces retain the input’s specific personality.
- Network architecture: The L2 reconstruction objective keeps generated faces similar to their inputs while training the encoder and generator.The objective is expressed as L(x, G(E(x), l)).
- Manifold traversal: The learned manifold represents age along one traversable direction and personality along another, enabling simultaneous age progression and regression.The framework differs from GANs, VAEs, and AAEs by combining encoder- and generator-side discriminators for smooth transitions and photo-realistic faces.
- Adversarial regularization: The encoder discriminator Dz imposes a uniform prior on z, filling latent-space holes and smoothing traversals between faces.A more evenly populated latent space is intended to support plausible intermediate morphing faces.
- Adversarial regularization: The image discriminator Dimg evaluates realism conditional on age, counteracting blurred or implausible outputs from pixel-wise reconstruction alone.It is reported to enhance texture, especially for older faces.
5. Experimental Evaluation
The evaluation uses a 10,670-image dataset and compares CAAE qualitatively and quantitatively with ground truth and prior face-aging methods. Results indicate realistic aging and rejuvenation, personality preservation, and robustness to pose, expression, and occlusion.
- Data Collection: The final dataset contains 10,670 face images distributed uniformly across ten age categories and gender.Images come from Morph, CACD, and web searches, with ages estimated from captions or an age estimator.
- Evaluation Setup: Testing on FGNET uses 1,002 images of 82 subjects spanning ages 0 to 69.The evaluation compares generated faces with ground truth and prior work.
- Ground-Truth Comparison: 48.38% of 3,208 votes judged generated faces the same person as the ground truth, versus 29.58% not and 22.04% unsure.The survey paired generated faces with ground-truth images whose age gap exceeded 20 years.
- Comparison with Prior Work: 52.77% of 1,508 votes preferred CAAE over prior methods, compared with 28.99% preferring prior work and 18.24% finding them equal.The comparison covered 235 paired images from 79 subjects and included both progression and regression comparisons.
- Qualitative Results: CAAE generates faces across age progression and regression while preserving personality and producing richer age-related texture over long age spans.The reported examples include wrinkles in older faces and higher fidelity than cited prior results.
- Robustness: CAAE produces generated faces for inputs with expression variation, non-frontal pose, and partial occlusion without removing those variations.The authors contrast this with prior approaches that often use face normalization and may still suffer from occlusion.
6. Discussion and Future Works
The discussion presents CAAE as a flexible framework for face-age-related tasks, with testing using only the encoder and generator. Its age-conditioned discriminator can also support age estimation, while the encoder preserves personal face features.
- Discussion: CAAE is presented as a holistic framework supporting both face age progression and regression.The framework learns a manifold rather than relying on conventional group-based training.
- Future Applications: At testing time, only E and G are used, while Dimg is trained conditionally on age to assess whether a face matches a specified age.The authors identify this discriminator role as directly corresponding to age estimation.
- Latent Representation: The encoder maps faces to latent vectors that preserve personality features for subsequent age-conditioned generation.This connects the latent representation to the framework’s stated goal of flexible age manipulation.