A half-day tutorial on RL-based and non-RL alignment methods for visual diffusion models.
Aligning diffusion models with human preferences has become an increasingly important research area in visual generation. This half-day tutorial gives a structured introduction to the area, organized around the four main families of alignment methods: direct reward backpropagation, which differentiates a reward through the sampling process; DPO-style preference optimization, which learns from pairs of preferred and rejected samples; GRPO-based online RL, which samples groups of outputs and learns from their relative rewards; and forward-process RL, which reformulates the alignment objective on the forward diffusion process, close to the pre-training objective.
For each family, we cover the core formulation, representative methods, and their strengths and limitations, and we close with open challenges and future directions.
