Hierarchical Variational Policies for Reward-Guided Diffusion
This work addresses the need for efficient test-time adaptation of diffusion models to downstream tasks, offering a practical speed-quality tradeoff for practitioners.
The paper proposes a hierarchical variational framework for adapting pretrained diffusion models to downstream objectives, achieving high-quality reward-aligned samples with significantly reduced inference cost. On 4x super-resolution, the method achieves better perceptual quality with more than 5x faster inference compared to the best-performing baseline.
Adapting pretrained diffusion models to downstream objectives such as inverse problems often requires expensive test-time guidance or optimization. We propose a principled framework for generating high-quality reward-aligned samples at substantially reduced inference cost. Our approach formulates test-time adaptation as a hierarchical variational model, where control is amortized into a lightweight yet expressive stochastic policy. This formulation naturally supports few-step diffusion sampling: large step sizes enable fast inference, while the learned policy maintains sample quality by providing structured per-step control. The resulting fully amortized sampler achieves a strong quality--speed tradeoff, matching or exceeding recent test-time scaling baselines while requiring significantly less compute. For example, on 4x super-resolution, our method achieves better perceptual quality with more than 5x faster inference compared to the best-performing baseline. We further extend our approach to a semi-amortized regime that combines cheap amortized proposals with limited test-time optimization, achieving state-of-the-art perceptual quality across several challenging inverse problems.