Efficient amplitude super-resolution for computer-generated holography via lightweight convolutional networks

Computer-generated holography for high-resolution displays demands computational efficiency alongside reconstruction fidelity. Existing approaches face a fundamental trade-off: high-resolution input processing achieves good fidelity but incurs prohibitive costs, while low-resolution methods suffer from insufficient amplitude detail. We present a systematic study of lightweight amplitude super-resolution architectures for holographic phase generation. Through controlled comparison of six efficient designs with shared downstream networks, we identify a compact architecture that achieves superior parameter efficiency. Our best model, employing depthwise separable convolution blocks with only ∼24K parameters, attains 32.66 dB PSNR on DIV2K validation and 32.77 dB on Flickr2K. Cross-dataset evaluation demonstrates consistent advantages in perceptual quality and parameter efficiency, offering a practical solution for holographic display systems.