LiftHolo: lifted amplitude-guided deep learning for accelerated high-quality computer-generated holography

Computer-generated holography (CGH) for high-resolution displays faces a critical trade-off: direct processing of high-resolution (HR) inputs achieves good fidelity but incurs a heavy computational burden, while low-resolution (LR) input methods suffer from insufficient amplitude information. We propose LiftHolo, a lightweight deep learning framework that performs amplitude super-resolution on LR inputs to guide complex-domain phase estimation. With only ∼44 K parameters, LiftHolo achieves a PSNR of 33.72 dB on DIV2K validation, outperforming the native HR-processing model by +1.67 dB while reducing computational complexity to ∼25%. Optical experiments on a UPOLabs HDSLM36R SLM further validate that LiftHolo maintains its reconstruction advantage in physical holographic display.