Phase retrieval reconstructs a complex-valued signal from intensity-only measurements. Modern learned regularizers are predominantly trained on real-valued datasets. Applying them to the magnitude and phase of phase images separately treats the two as independent and can miss their physical correlation. The project aims to train a denoiser that operates directly on complex numbers, regularizing magnitude and phase jointly. The candidate will first study how amplitude and phase are coupled across imaging modalities and use this to construct complex-valued datasets. A complex-valued denoiser will then be trained and integrated into reconstruction, comparing several paradigms — Plug-and-Play, diffusion, and flow-matching — and the associated design choices. The project concludes with a benchmark quantifying whether the joint complex denoiser outperforms magnitude/phase splitting.