Untrained physics-enhanced fully complex transformer for single-frame hologram reconstruction

Single-frame hologram reconstruction without training remains challenging for real-world samples due to the difficulty of modeling coupled amplitude and phase information. To address this issue, we develop a fully complex-domain attention mechanism that reformulates attention computation directly in the complex field using complex inner products, enabling more effective representation of amplitude-phase coupling. Based on this formulation, a general fully complex-domain transformer, termed HoloFormer v3, is constructed and instantiated within an untrained reconstruction framework. An auxiliary free-space back-propagation initialization reformulates the highly nonlinear hologram-to-object complex amplitude inversion as a twin-image-suppressed optimization in the object plane. Experiments on real samples demonstrate improved reconstruction accuracy and faster convergence compared with existing untrained methods.