Understanding and controlling facial attributes with StyleGAN

Song, Shuang 2025. Understanding and controlling facial attributes with StyleGAN. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF (Thesis) - Accepted Post-Print Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (30MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (221kB)

Abstract

Facial attributes such as age, gender and accessories are fundamental cues in human perception and central to recognition, retrieval and controllable image manipulation. Recent advances in generative models, particularly, GAN-based approaches have enabled high-fidelity facial synthesis with increasing levels of semantic control. However, their internal mechanisms remain only partially understood. In particular, GAN still exhibits unstable visual patterns and incomplete semantic disentanglement in its latent and feature representations, constraining both representation learning and reliable attribute manipulation. This thesis addresses these limitations by treating StyleGAN’s internal feature maps as a principled basis for improving generation stability, unsupervised attribute representation and controllable manipulation. First, we identify a previously overlooked architectural issue feature proliferation in which intermediate activations become excessively amplified and produce characteristic artifacts. We introduce a lightweight feature-rescaling regularization that suppresses this proliferation and substantially stabilises StyleGAN generation while preserving sample diversity. Building on this improved stability, we show that StyleGAN’s intermediate feature maps encode richer and more separable attribute semantics than conventional latent spaces. We develop an unsupervised representation learning framework that extracts compact semantic descriptors from these activations, enabling effective clustering and attribute annotation and achieving competitive or superior performance compared with strong self-supervised baselines. Finally, we analyse how latent style parameters interact with feature maps to form semantic attributes, revealing why style-only manipulation often fails to produce precise and reliable control. Based on these insights, we propose a joint style–feature manipulation framework that strengthens relevant channels and aligns them with style modulation, achieving more precise, robust and identity-preserving manipulation across multiple facial attributes. Overall, this thesis advances the understanding of StyleGAN’s internal mechanisms and establishes feature maps as a unifying foundation for more stable generation, stronger unsupervised semantic representation and more accurate controllable manipulation.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Funders:	China Scholarship Council
Date of First Compliant Deposit:	5 June 2026
Date of Acceptance:	28 May 2026
Last Modified:	05 Jun 2026 12:14
URI:	https://orca.cardiff.ac.uk/id/eprint/187416

Actions (repository staff only)

Edit Item