Hubness awareness sampling for deep generative models in generation and evaluation

Liang, Yuanbang ORCID: https://orcid.org/0009-0000-8370-6655 2025. Hubness awareness sampling for deep generative models in generation and evaluation. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (49MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (202kB) | Request a copy

Abstract

Despite the rapid progress in Generative Adversarial Networks (GANs), several fundamental challenges remain under-explored, including reliable latent sampling, scalable evaluation, and fairness in generation. In this work, we propose a unified framework based on hubness sampling, a principle derived from the observation that high-dimensional latent spaces exhibit hub latents. We show that these hub latents are better trained and contribute more to the synthesis of high-quality images. Leveraging this insight, we develop an a priori latent sampling method that outperforms traditional approaches such as the empirical truncation trick, both in efficiency and image quality. Building on this foundation, we address the computational bottlenecks in eval uating generative models on large datasets. We introduce efficient precision and recall (eP&R) metrics that retain fidelity to the original metrics while significantly reducing computation through hubness-aware sampling and approximate nearest neighbor techniques. Finally, we extend hubness sampling to promote fairness and diversity in GAN training. Without requiring labels or additional supervision, hubness sampling improves representation across sensitive attributes such as ethnicity, gender, and age, applied to various state-of-the-art GAN architectures, including StyleGAN, Diffusion-GAN, and GANFormer. In conclusion, this work demonstrates that hubness sampling offers a versatile and powerful toolset for improving image quality, evaluation efficiency, and fairness in generative modeling, while also highlighting opportunities for further optimization in its computational cost.

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Status:	Unpublished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:	EPSRC DTP studentship
Date of First Compliant Deposit:	4 February 2026
Date of Acceptance:	7 December 2025
Last Modified:	04 Feb 2026 14:23
URI:	https://orca.cardiff.ac.uk/id/eprint/184406

Actions (repository staff only)

Edit Item