Abstract
Diffusion models have emerged as the state-of-the-art in generative modeling, producing remarkably high-fidelity images, audio, and more. Despite their success, two major hurdles often limit their practical deployment: the slow, iterative nature of their sampling process and the challenge of aligning their powerful generative capabilities with nuanced human intent.

This talk addresses these two critical challenges in two parts. First, I will present our work on enhancing the controllability of diffusion models by incorporating human feedback directly into the fine-tuning process. We explore how reinforcement learning and preference-based methods can steer generation to better match subjective and complex user intentions, moving beyond the limitations of text-based prompts.

In the second part, I will tackle the issue of sampling efficiency by introducing a novel framework for learning a generative flow map. By reformulating the diffusion process as a direct, one-step mapping from noise to data, our approach enables high-quality synthesis in a single network evaluation. This method, which stems from the principles of diffusion models, significantly accelerates generation without compromising quality. Together, these advancements pave the way for a new generation of diffusion models that are not only powerful but also efficient, controllable, and collaborative.