Face Aging via Diffusion-based Editing

Xiangyi Chen, Stéphane Lathuilière

Introduction

Face aging, also known as facial age progression and regression, involves generating past or future versions of a person's face while maintaining their unique identity. Traditionally, this task has been dominated by Generative Adversarial Networks (GANs). However, GAN-based methods suffer from two primary limitations: the data bottleneck and the inversion gap. Because GANs are typically trained on specific, often biased datasets like FFHQ or CelebA, they struggle with extreme age gaps and rare facial orientations. Furthermore, mapping a real-world photo into a GAN's latent space often results in a loss of identity or the introduction of artifacts.

To address these challenges, we propose FADING (Face Aging via DIffusion-based editiNG). By leveraging the massive visual priors of large-scale text-to-image diffusion models, FADING overcomes the constraints of traditional datasets and provides a more robust framework for high-fidelity face aging across the entire human lifespan.

The FADING Framework

The FADING pipeline utilizes a three-stage process to transform facial images accurately while preserving identity.

1. Age-Aware Fine-Tuning

Rather than using a generic pre-trained model, we specialize a Latent Diffusion Model (LDM) for the aging task. We employ a "double-prompt" fine-tuning scheme, training the model with pairs of prompts: a specific prompt (e.g., "photo of a [age] year old person") and a generic prompt ("photo of a person"). This strategy allows the model to disentangle age-related features from identity-related features, learning that age is a variable attribute while the core "person-ness" remains constant.

2. Null-Text Inversion

To edit a real-world photograph, the model must first be able to reconstruct it perfectly. We utilize Null-Text Inversion, which optimizes the unconditional text embeddings. This ensures that the diffusion process can recreate the input image with near-perfect fidelity before any age-related modifications are applied.

3. Localized Editing via Attention Control

During the actual aging process, we swap the original age token for a target age (e.g., changing "30" to "70"). To ensure the structural integrity of the face remains unchanged, we implement attention control. By injecting the spatial attention maps from the original image into the new generation process, we ensure that age-related changes—such as wrinkles or graying hair—appear in the correct anatomical locations without altering the background, clothing, or fundamental face shape.

Experimental Evaluation

We evaluated FADING using two high-resolution benchmarks:

  • FFHQ-Aging: An extension of the NVIDIA FFHQ dataset containing 70,000 images across 10 age groups.
  • CelebA-HQ: A dataset of 30,000 celebrity images used to test the model's ability to generalize across different identities.

Quantitative Results

FADING demonstrates significant improvements over previous state-of-the-art (SOTA) methods like CUSP and HRFAE:

  • Aging Accuracy: On the FFHQ-Aging dataset, FADING achieved a Mean Absolute Error (MAE) of 9.16, compared to 12.63 for CUSP.
  • Identity Preservation: The model maintained a 98.4% gender preservation rate on CelebA-HQ, indicating that fundamental identity attributes are not compromised during the aging process.
  • Visual Quality: Using the Kernel-Inception Distance (KID) metric, FADING showed nearly an order of magnitude improvement in image realism across most age groups.

Discussion and Insights

The primary advantage of FADING lies in its use of the "Internet Prior." Unlike GANs trained on tens of thousands of faces, diffusion models trained on billions of images possess a deep understanding of how aging affects diverse ethnicities, lighting conditions, and poses. This makes FADING exceptionally robust to "rare" cases, such as faces with sunglasses, hats, or extreme head tilts, which often cause GANs to hallucinate or distort features.

However, we did observe a slight entanglement between age and skin tone, likely a bias inherited from the underlying Stable Diffusion training data. Future work will focus on enhancing the prompt strategy to include age-agnostic attributes—such as "wearing glasses"—to ensure these features are perfectly preserved or realistically aged alongside the face.

Conclusion

FADING represents a significant step forward in digital face aging. By repurposing large-scale diffusion models and employing sophisticated inversion and attention control techniques, we achieve superior aging accuracy and identity preservation compared to traditional GAN-based approaches. This framework opens new possibilities for realistic facial transformations across the human lifespan.

Summary updated

Abstract

In this paper, we address the problem of face aging: generating past or future facial images by incorporating age-related changes to the given face. Previous aging methods rely solely on human facial image datasets and are thus constrained by their inherent scale and bias. This restricts their application to a limited generatable age range and the inability to handle large age gaps. We propose FADING, a novel approach to address Face Aging via DIffusion-based editiNG. We go beyond existing methods by leveraging the rich prior of large-scale language-image diffusion models. First, we specialize a pre-trained diffusion model for the task of face age editing by using an age-aware fine-tuning scheme. Next, we invert the input image to latent noise and obtain optimized null text embeddings. Finally, we perform text-guided local age editing via attention control. The quantitative and qualitative analyses demonstrate that our method outperforms existing approaches with respect to aging accuracy, attribute preservation, and aging quality.
BibTeX
@inproceedings{chen2023face,
  author = {Xiangyi Chen and Stéphane Lathuilière},
  title = {{Face Aging via Diffusion-based Editing}},
  booktitle = {BMVC},
  year = {2023},
}
Figure