Skip to main content
eScholarship
Open Access Publications from the University of California

UC San Diego

UC San Diego Electronic Theses and Dissertations bannerUC San Diego

Debiasing Image Generative Models

Abstract

Generative models have become increasingly popular in various domains to solve challenging tasks, including image generation, dialogue generation, and story generation. Unlike discriminative models, they can learn the underlying probability distribution of data and generate new examples. In particular, image generative models have gained significant attention due to their remarkable ability to produce images of unparalleled quality. However, while there has been a lot of attention to biases in discriminative models, bias in generative models has received little attention. The presence of biases in generative models, particularly related to race and gender, can have significant consequences in downstream applications. Therefore, efforts to address this issue are essential to promote fair and ethical use of generative models in various domains. To achieve this goal, this dissertation presents a comprehensive study of debiasing image generative models by incorporating diversity and fairness constraints into the training process.

In this dissertation, we investigate three different approaches to debiasing image generative models. In the first approach, a new task of high-fidelity image generation conditioned on multiple attributes from imbalanced datasets is proposed. This task poses new challenges for state-of-the-art GANs models, and a new training framework is proposed to address thesechallenges. The second approach investigates bias in image-to-image translation models and proposes debiasing using contrastive learning. Finally, the study highlights the prevalence of bias in large-pretrained models like CLIP and its impact on text-to-image generative models. Identity preserving losses are proposed to rectify the problem without retraining the pretrained model. In all of these approaches, we evaluate the impact of debiasing on image generation and the effectiveness of existing methods in reducing biases in generated images. We show the proposed task and framework offer new avenues for further research in debiasing generative models. Overall, this dissertation contributes to the field of generative models by providing a comprehensive study of debiasing generative models and proposing a new task and framework for high-fidelity image generation.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View