Swift hash-convolutional neural representation for differentiable computational imaging

Abstract Computational imaging solves inverse problems by combining physical models with algorithmic reconstruction. Conventional methods rely on discrete pixel grids, which fix resolution, consume extensive memory, and fail to capture the continuous nature of optical fields. Implicit neural representations (INRs) offer a continuous alternative but are limited by the slow convergence and computational inefficiency of multi-layer perceptron (MLP) decoders. Here, we introduce the swift hash-convolutional neural representation (SWAN), a computational framework that overcomes the above limitations by integrating a multi-resolution hash encoding with a lightweight convolutional decoder. The hash encoding stores multi-scale spatial features compactly in learnable embedding tables, while the convolutional decoder explicitly models local spatial correlations-shifting the paradigm from inefficient pixel-by-pixel prediction to a spatially coherent, physics-aligned representation. When embedded into physics-driven models for lensless holography, coded ptychography and coded illumination ptychography, SWAN achieves faster convergence, higher reconstruction fidelity, and superior memory efficiency compared to both conventional iterative algorithms and MLP-based INRs. Without requiring pre-training, SWAN adapts directly to raw measurements, establishing a flexible, high-performance backbone for next-generation computational imaging systems.