This paper focuses on diffusion/score-based models, which have recently emerged as powerful generative dictionaries for solving inverse problems, particularly for accelerated MRI reconstruction. While the flexibility of these models allows for the separation of measurement models and learned dictionaries, their performance relies heavily on carefully tuned data fidelity weights, particularly under fast sampling schedules with few denoising steps. Existing approaches often rely on heuristics or fixed weights, failing to generalize to diverse measurement conditions and irregular time-step schedules. In this study, we propose Zero-shot Adaptive Diffusion Sampling (ZADS), a test-time optimization method that adaptively adjusts fidelity weights under arbitrary noise schedules without retraining the diffusion dictionary. ZADS treats the denoising process as a fixed, unrolled sampler and optimizes fidelity weights in a self-supervised manner using only undersampled measurements. Experiments on the fastMRI knee dataset demonstrate that ZADS consistently outperforms existing compressed sensing and recent diffusion-based methods, providing high-fidelity reconstruction across a variety of noise schedules and acquisition settings.
