Face aging is an ill-posed problem because multiple plausible aging patterns may correspond to a given input. Most existing methods often produce one deterministic estimation. This paper proposes a novel CLIP-driven Pluralistic Aging Diffusion Autoencoder (PADA) to enhance the diversity of aging patterns. First, we employ diffusion models to generate diverse low-level aging details via a sequential denoising reverse process. Second, we present Probabilistic Aging Embedding (PAE) to capture diverse high-level aging patterns, which represents age information as probabilistic distributions in the common CLIP latent space. A text-guided KL-divergence loss is designed to guide this learning. Our method can achieve pluralistic face aging conditioned on open-world aging texts and arbitrary unseen face images. Qualitative and quantitative experiments demonstrate that our method can generate more diverse and high-quality plausible aging results.
Official Implementation of our Pluralistic Aging Diffusion Autoencoder (PADA) paper for both training and evaluation. PADA allows modeling flexible and diverse face aging conditioned on texts and images.
- Release the pretrained model and inference code.
- Release the training code.
- Add jupyter notebooks demo.
Our code borrows from SAM and DiffAE. We would like to express our gratitude for their generosity in sharing their work.
If you use this code for your research, please cite our paper Pluralistic Aging Diffusion Autoencoder:
@InProceedings{Li_2023_ICCV,
author = {Li, Peipei and Wang, Rui and Huang, Huaibo and He, Ran and He, Zhaofeng},
title = {Pluralistic Aging Diffusion Autoencoder},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2023},
pages = {22613-22623}
}