More about HKUST
Diffusion-based Controllable Image Generation: Methods, Applications and Future Directions
PhD Qualifying Examination
Title: "Diffusion-based Controllable Image Generation: Methods, Applications
and Future Directions"
by
Mr. Yanghao WANG
Abstract:
With the surge in image generation of recent years, the high-fidelity image
generation problem has been well addressed by state-of-the-art generative
models, i.e., diffusion models. However, humans' generation requirements have
also exploded, which requires not only fidelity but also controllability,
that is, controllable image generation. Specifically, controllable image
generation leverages control signals to drive the generation process and aims
to obtain regularized results that follow these signals. To better understand
the context of controllable image generation, we focus on two aspects: 1) How
to represent the control signals and inject them into the vanilla diffusion
models. 2) How to unleash the capacity of controllable image generation for
various applications. In this survey, we provide a comprehensive analysis of
diffusion-based controllable image generation. First, we analyze its control
methods and summarize them into nine mechanisms across three stages.
Moreover, we demonstrate its abundant applications in seven scenarios.
Finally, we discuss the existing challenges and the further extensions.
Date: Monday, 17 November 2025
Time: 4:00pm - 6:00pm
Venue: Room 5566
Lifts 27/28
Committee Members: Dr. Long Chen (Supervisor)
Prof. James Kwok (Chairperson)
Dr. Dan Xu