Hierarchical Variational Policies Boost Diffusion Model Speed

A new framework from arXiv proposes hierarchical variational policies to adapt pretrained diffusion models for downstream tasks like inverse problems at reduced inference cost. The method amortizes control into a lightweight stochastic policy, enabling few-step sampling with large step sizes for fast inference while maintaining quality. On 4x super-resolution, it achieves better perceptual quality with over 5x faster inference compared to test-time scaling baselines.