Unsupervised Learning of Facial Attribute Representations Using StyleGAN

shuang song, Jing Wu, Yu-Kun Lai, Yipeng Qin

Keywords: facial attributes, unsupervised representation learning, GAN, StyleGAN

Abstract:

Facial attributes (e.g., gender, age) encompass important social cues and play a pivotal role in computer vision. While supervised methods have dominated facial attribute analysis, they often require large annotated datasets, which are costly and time-consuming to create.
In this work, we circumvent this limitation by proposing a novel unsupervised learning framework that leverages StyleGAN to learn rich and disentangled facial attribute representations. Specifically, unlike prior methods that rely on labeled datasets or supervised techniques, our approach exploits the unique inductive bias of StyleGAN, namely Hierarchical Feature Modulation, to automatically discover semantically meaningful representations of facial attributes. This inductive bias enables StyleGAN to generate disentangled and interpretable facial attribute features at different layers, benefiting a variety of downstream tasks. To leverage StyleGAN representations, we employ GAN inversion methods to represent input images as StyleGAN features and propose a simple yet effective feature reduction method based on mutual information to improve the effectiveness and efficiency of the learned representations. Extensive experiments in few-shot facial attribute analysis tasks, including clustering, classification, and facial attribute annotation demonstrate the effectiveness of our approach.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9940