Researchers Introduce DM-Align for Video Generation Models

Researchers have introduced DM-Align, a unified single-stage optimization framework designed to align and distill video generative models simultaneously. The findings were detailed in a preprint published on arXiv.

Aligning video generation models with human preferences typically relies on reinforcement learning (RL) coupled with distillation in separate stages. However, executing RL before distillation is computationally prohibitive, while applying it afterward often triggers model collapse.

DM-Align resolves these challenges by applying sample-guided distribution matching. Drawing inspiration from DPO and GRPO, the framework avoids multi-step reward evaluations and complex conversions, improving distillation quality and preference alignment across foundational video models.