通过解耦频段学习，用扩散模型修复相位图细节

Oversmoothing remains a persistent problem when applying deep learning to off-axis quantitative phase imaging (QPI). End-to-end U-Nets favour low-frequency content and under-represent fine, diagnostic detail. We trace this issue to spectral bias and show that the bias is reinforced by high-level skip connections that feed high-frequency features directly into the decoder. Removing those deepest skips thus supervising the network only at a low resolution significantly improves generalisation and fidelity. Building on this insight, we introduce DiffPR, a two-stage frequency-decoupled framework. Stage 1: an asymmetric U-Net with cancelled high-frequency skips predicts a quarter-scale phase map from the interferogram, capturing reliable low-frequency structure while avoiding spectral bias. Stage 2: the upsampled prediction, lightly perturbed with Gaussian noise, is refined by an unconditional diffusion model that iteratively recovers the missing high-frequency residuals through reverse denoising. Experiments on four QPI datasets (B-Cell, WBC, HeLa, 3T3) show that DiffPR outperforms strong U-Net baselines, boosting PSNR by up to 1.1 dB and reducing MAE by 11 percent, while delivering markedly sharper membrane ridges and speckle patterns. The results demonstrate that cancelling high-level skips and delegating detail synthesis to a diffusion prior is an effective remedy for the spectral bias that limits conventional phase-retrieval networks.