TL;DR
This paper introduces TimeBooth, a novel approach to personalized face aging that moves beyond traditional numerical age representations. By modeling age as an "Age Prompt" (an image-based representation) and disentangling identity and age features through adversarial learning, TimeBooth generates more diverse and controllable aged faces. This work establishes a new paradigm for face aging by leveraging the richness of image-based age priors.
What Problem Does This Paper Address?
Face aging, the process of synthesizing photorealistic faces at different ages while preserving identity, has numerous applications in fields like film, forensics, and diagnostics. However, existing methods often struggle with the inherent ill-posed nature of face aging, which is influenced by a complex interplay of genetics, environment, and lifestyle. These methods typically rely on numerical age labels, which are too simplistic to capture the diverse aging patterns observed in real individuals. This leads to generated images that often reflect an "average" aging process rather than personalized outcomes, lacking the nuanced variations that make aging unique to each person. The core challenge is to represent age in a way that is rich enough to guide the generation of diverse and individualized aged faces.
Methods
TimeBooth tackles this challenge with a multi-pronged approach:
- Disentangled Facial Representation: The core innovation is representing age not as a number, but as an "Age Prompt" – an image that visually embodies the desired age characteristics. This is achieved by learning disentangled representations for identity (ID) and age.
- ID-Age Representation Co-learning Framework: A multi-task learning framework is employed, using ID recognition and age classification tasks to jointly supervise the learning of these disentangled features.
- Bidirectional Adversarial Disentanglement (BAD) Strategy: To ensure that ID and age representations are truly independent, a bidirectional adversarial learning approach with mutual Gradient Reversal Layers (GRLs) is used. This forces the model to learn attribute-invariant features, meaning the ID representation remains stable across different ages, and vice-versa.
- Age Adapter: A lightweight Q-former is used to fuse the disentangled ID and age features. This fused representation is then fed into a diffusion model via cross-attention layers to generate the aged face.
- Training Strategies: To optimize the Age Adapter and ensure robust learning, several strategies are introduced:
- Shuffle Strategy: Dynamically replaces ID features with those of the same identity at different ages during training to improve generalization.
- Age Prototype Replacement (APR) Strategy: Replaces individual age features with cohort prototypes to leverage population-level aging patterns and prevent spurious correlations.
- Dynamic Training-set Sampling (DTS) Strategy: Addresses age distribution imbalance in datasets by adaptively selecting samples.
- Personalized Aging during Inference: A Spherical Linear Interpolation method is used to fuse the source age feature with the target Age Prompt, controlled by an aging guidance parameter ($\alpha$) to balance editability and fidelity.
Datasets
- MS1Mv2: Used in the first training stage for ID recognition and age estimation.
- Hybrid Dataset: Constructed for the second stage, comprising approximately 500K high-quality facial images. This dataset includes:
- A general cross-age dataset without identity overlap (FFHQ and SFHQ).
- A cross-age dataset with the same identity across different ages (derived from VGGFace2-HQ) to create an identity pool.
Key Insights and Ideas
- Image as Age Prior: The most significant insight is the shift from numerical age labels to image-based "Age Prompts." This allows the model to leverage the rich, semantic information present in images to guide face aging, leading to more nuanced and realistic results.
- Disentanglement is Key: The success hinges on effectively disentangling identity from age. The BAD strategy, using bidirectional adversarial learning, is crucial for achieving this, ensuring that identity is preserved while age is transformed.
- Data-driven Aging Patterns: The Age Prompt approach, combined with strategies like APR, allows the model to learn and generate diverse aging patterns directly from data, moving beyond predefined rules.
- Plug-and-Play Efficiency: TimeBooth's design, particularly its Age Adapter, allows for efficient, tuning-free integration into existing diffusion models, making it highly practical.
Key Results
- Qualitative Superiority: TimeBooth generates more natural and diverse aged faces compared to translation-based methods (CUSP, IP2P, FADING) and baseline methods (DreamBooth, FaceID). It effectively avoids the "child faces with elderly textures" issue and produces results with semantics better aligned with real aging.
- Quantitative Performance: In Table 1, TimeBooth achieves superior performance across several metrics, including KID (lower is better), Face-sim (higher is better), AgeMAE (lower is better), and R-AgeMAE (lower is better). Notably, it outperforms baselines on most metrics.
- User Study: A user study confirmed TimeBooth's superiority in facial consistency, Age Prompt similarity, and visual quality compared to baselines.
- Inference Speed: TimeBooth achieves fast inference times (<10s), significantly outperforming methods like DreamBooth which require fine-tuning and take much longer.
- Aging Guidance Control: The aging guidance parameter ($\alpha$) effectively controls the degree to which Age Prompt features influence the generated image, allowing for fine-tuned control over the aging process.
Limitations and Future Work
- Limitations: The paper acknowledges that while translation-based methods might offer higher fidelity in some aspects, TimeBooth's strength lies in its diversity and controllability. The "child faces with elderly textures" issue, while mitigated, might still be a challenge for some extreme transformations.
- Future Work: The authors suggest exploring invariant face representations further and extending their applications to broader and more general face generation tasks. This implies a potential for applying the disentanglement and Age Prompt concepts to other attribute manipulation tasks beyond aging.
Summary updated