This talk studies diffusion models across the pipeline from pretraining to post-training and posterior sampling. We first analyze how the forward diffusion schedule affects generation accuracy and sampling complexity, providing guidance for the design of stronger base models. Given a pretrained model, we then consider two ways to steer its generative distribution toward specific objectives: reinforcement-learning post-training for reward alignment and direct posterior sampling for inverse problems. Together, these topics illustrate how diffusion models can be designed, analyzed, and controlled from base-model construction to downstream applications.
