Snapshot compressive imaging reconstruction: from traditional AI to generative AI

Snapshot compressive imaging (SCI) overcomes a fundamental bottleneck of traditional imaging by capturing multi-dimensional data (e.g., spectral or temporal) with a single 2D sensor exposure. The past decade has seen the transition of SCI from theory to real-world applications—such as hyperspectral imaging and highspeed video—largely driven by artificial intelligence (AI) that enables rapid, high-fidelity reconstruction. While reconstruction backbones have evolved from CNNs to Transformers and state-space models (e.g., Mamba), these approaches can be broadly classified as traditional AI. Recently, generative AI, particularly diffusion models, has emerged as a powerful paradigm due to its strong prior modeling and synthesis capability. Leveraging such generative models for SCI reconstruction is a natural yet challenging step, primarily due to the domain gap between the encoded measurements and realistic scene generation. This paper reviews recent advances in generative AI for SCI reconstruction and introduces a novel framework designed for both spectral and video SCI.