UCLA

The rapid advancement of generative models, particularly in image generation and manipulation, has opened new possibilities in creative and design fields. However, their complex 'black box' nature poses significant challenges for user interaction and control, especially for non-experts. This dissertation addresses these challenges by developing and evaluating user-centric tools that enhance interaction with Generative Adversarial Networks (GANs) and Text-to-Image (T2I) models.

Central to this work is the exploration of user-driven methods to improve the usability and accessibility of these models. The research introduces innovative tools that empower users to iteratively refine and control the generative process. These tools are designed to complement existing GAN architectures, allowing users to interact more intuitively with the models, particularly in tasks requiring precise image editing and creative content generation.

Empirical studies form a significant part of this research, evaluating the effectiveness of these tools in real-world scenarios. The studies involve user tasks in image editing and creative content generation. Findings demonstrate that the developed tools not only facilitate a more intuitive interaction with generative models but also enable users to achieve superior results compared to existing state-of-the-art methods.

Additionally, this dissertation investigates ways to enhance user understanding of the generative process. By making the mechanisms of these models more transparent and comprehensible, the research contributes to a more informed and effective use of these technologies.

In conclusion, this dissertation focuses on making advanced generative models more accessible and user-friendly. It offers insights into the development of intuitive tools that bridge the gap between the complex capabilities of generative AI models and the creative and practical needs of users.