Modern imaging technologies are central to progress in science, medicine, and engineering. Yet, many advanced imaging systems operate under physical or resource limitations that make it difficult to directly acquire high-quality images. Computational imaging addresses these challenges by using algorithms to reconstruct images from incomplete or indirect measurements. In recent years, deep learning has enabled new capabilities in computational imaging, but current methods assume that the training and test data share the same conditions. This assumption often does not hold in real-world settings. This project addresses this critical gap by developing new methods to ensure that deep-learning models for image reconstruction remain reliable and accurate even when the data conditions shift. The outcomes of this research will have broad use and transformative effects across a wide range of scientific, engineering, and biomedical applications, where robust image reconstruction is essential. Broader-impact activities include the organization of special sessions, workshops, and journal issues for the computational-imaging community, dissemination via open-source code, and curriculum development at both institutions. This project focuses on score-based models — a class of deep generative models that solve imaging problems by learning the score function of the image distribution. The central goal is to develop a unified mathematical framework for analyzing and improving the robustness