Text-Guided Diffusion Model for Adaptive Reconstruction through Scattering Media

Optical imaging through scattering media presents a fundamentally ill-posed inverse problem, as inhomogeneous materials scramble incident light into seemingly random speckle patterns, destroying phase information and diffraction-limited resolution. While the optical memory effect allows for the recovery of object autocorrelation from scattered intensity, traditional phase-retrieval algorithms routinely stagnate on complex or dense objects. In this paper, we introduce a text-guided diffusion model framework to solve the inverse problem of imaging through scattering layers, leveraging the strong generative priors of diffusion models and the semantic disambiguation of text conditioning. We evaluate our method on the MNIST and Quick, Draw! datasets using experimentally obtained speckle measurements through an optical diffuser. Our results demonstrate that textual prompts effectively resolve the inherent ambiguities of phase loss, significantly outperforming traditional correlation-based imaging techniques. As a foundational proof of concept, this work isolates the phase-retrieval reconstruction step to demonstrate that explicit semantic priors can resolve topological ambiguities where classical methods fail.