Dissertation
Visual generation systems must follow user intent expressed through prompts, structural controls, editing instructions, and preferences. Optimizing a standard denoising loss alone does not directly ensure that the generated result satisfies these goals.
This dissertation develops a reward-centered approach to visual alignment. ControlNet++ turns condition-output consistency into a training signal. SuperEdit repairs inaccurate editing supervision. Semi-DPO separates consistent preferences from conflicting ones and refines noisy labels. ViPO studies robust preference optimization and balanced preference data at scale.
Together, these contributions address the design, correction, relabeling, and scaling of reward signals. They provide methods for more controllable generation, better instruction-based editing, and more reliable preference alignment in image and video generation.