Scene‐Customized Learning for Multi‐Depth 3D Phase‐Only Hologram Generation

Convolutional neural network (CNN)–based computer‐generated holography (CGH) enables high‐speed, high‐quality holographic display, yet its performance depends critically on the statistical characteristics of training datasets, and systematic frameworks for evaluating the performance boundaries of CNN‐based hologram encoders remain lacking. From a dataset design perspective, this study proposes GM‐4K, a 4K RGB‐D dataset constructed through scene‐customized geometric modeling. Using random geometric primitives, procedural textures, and parameterized spatial sampling, GM‐4K generates scenes with controllable low‐, mid‐, high‐, and wide‐frequency intensity distributions, while also allowing flexible depth‐region sampling. Numerical and optical experiments using a U‐Net++ encoder demonstrate that the intensity spectral distribution of training data significantly influences the reconstruction quality of multi‐depth phase‐only holograms and that models trained on wide‐ or mid‐frequency datasets exhibit better overall generalization in complex 3D scenes. Based on this observation, a spectral test framework comprising low‐, mid‐, and high‐frequency scenes is developed to evaluate hologram encoding models under different frequency conditions. The proposed dataset construction paradigm and spectral test framework provide controllable data support for multi‐depth hologram generation and offer practical guidance for data design and model evaluation in learning‐based CGH.