Diverse and Lifespan Facial Age Transformation Synthesis with Identity Variation Rationality Metric

TL;DR

This work tackles diverse and lifelike facial aging across the entire lifespan while preserving subject identity. It introduces a dual-path framework, DLAT+ (DLAT_{img} for texture and DLAT_{lmk} for geometry), with image warping to fuse results and a novel IDAG metric to quantify rational identity changes over age gaps. The approach demonstrates superior diversity, perceptual realism, and identity-consistency across long age spans, outperforming state-of-the-art lifespan-aging methods and offering a principled metric for identity rationality. The methodology enables robust data augmentation and downstream applications, while acknowledging limitations and avenues for future enhancements such as 3D priors and diffusion-based generative models.

Abstract

Face aging has received continuous research attention over the past two decades. Although previous works on this topic have achieved impressive success, two longstanding problems remain unsettled: 1) generating diverse and plausible facial aging patterns at the target age stage; 2) measuring the rationality of identity variation between the original portrait and its syntheses with age progression or regression. In this paper, we introduce ${\rm{DLAT}}^{\boldsymbol{+}}$ to realize Diverse and Lifespan Age Transformation on human faces, where the diversity jointly manifests in the transformation of facial textures and shapes. Apart from the diversity mechanism embedded in the model, multiple consistency restrictions are leveraged to keep it away from counterfactual aging syntheses. Moreover, we propose a new metric to assess the rationality of Identity Deviation under Age Gaps (IDAG) between the input face and its series of age-transformed generations, which is based on statistical laws summarized from plenty of genuine face-aging data. Extensive experimental results demonstrate the uniqueness and effectiveness of our method in synthesizing diverse and perceptually reasonable faces across the whole lifetime.

Figures (12)

• Figure 1: An overview of our method. During the training stage, the proposed $\rm{DLAT}_{img}$ (a) and $\rm{DLAT}_{lmk}$ (b) are learned separately. While the former learns to generate multiple age-transformed faces under each specific age condition with diversity mainly manifesting in the facial texture, the latter purely focuses on synthesizing structural variety on faces represented by landmarks. Regarding the $\rm{DLAT}_{img}$ model, its age mapping module $M_{img}$ is used to learn the mapping from a set of randomly sampled noise to $K$ sets of age latent codes correlated with $K$ different age stages. Moreover, those codes under each specific age phase need to be distinctive from each other. This goal is achieved with the help of other three components. Concretely, a face generator $G_{img}$ uses the source portrait and conditions on age latent codes to generate new faces located in specified age stages, which is ensured by a face discriminator $D_{img}$ with an adversarial loss $\mathcal{L}_{adv}^{img}$. Given these converted faces, a predictor $P_{img}$ is leveraged to estimate corresponding age latent codes, the conformance between which and the original ones is enforced by the $\mathcal{L}_{age}^{img}$. For those generated faces within the same age group, they are encouraged by $\mathcal{L}_{div}^{img}$ to be texturally different. To avoid pursuing unreasonable diversity, consistency of identity and race attributes are respectively required by $\mathcal{L}_{idc}$ and $\mathcal{L}_{rac}$ between generations and the source face. On the other hand, the age latent code predicted from the source image accompanied with age-shifted face syntheses are sent to the $G_{img}$ again with the supervision of $\mathcal{L}_{cyc}^{img}$ and $\mathcal{L}_{ppc}$, aiming at the reconstruction purpose. At the same time, similar diversity mechanisms and constraints designed for eliminating irrational transformation of facial geometry are imposed on the learning of $\rm{DLAT}_{lmk}$. After that procedure, pre-trained age mapping modules $M_{img}$ and $M_{lmk}$, as well as relevant generators $G_{img}$ and $G_{lmk}$, are picked out and integrated via image warping operations to form the $\rm{DLAT}^{\boldsymbol{+}}$ (c). In brief, it is an enhanced model capable of simulating age transformation effects across the life cycle on the given input faces with plausible variations both in facial appearance and shapes.
• Figure 2: Examples of poor age transformation synthesis that apparently contradict universal aging rules on human faces. However, these generations are appreciated when simply considering high face verification rates indicate more reasonable preservation of identity information across ages. The verification rates are $91.27\%$ between (a) and (b), $91.58\%$ between (a) and (c), measured by Face++ API FACEplus:API.
• Figure 3: Statistical analysis on real face-aging datasets about average variation trends of identity information across ages. For each portrait of a specific age condition (the source age stage), we take it as the anchor and compare its identity similarities with faces from the same subject at identical or different phases. Similarity scores are calculated via the Face++ API. The symbols A to J denote ten consecutive age groups. These figures clearly show that identity loss becomes more severe as the increase of gap between the source and target age stages.
• Figure 4: Visual comparison among lifespan age transformation results of the given male subject, which are synthesized by LATS, DLFS, SAM, CUSP, AgeTransGAN, InterFaceGAN, Latent-Transformer, and our $\rm{DLAT}^{\boldsymbol{+}}$.
• Figure 5: Visualization of diverse age transformation results across the life cycle, generated by applying our full method to a female subject.