Mitigating forward model mismatch in inverse problems via learned residuals and diffusion priors

Computational imaging, which solves inverse problems through classical, data-driven, or hybrid approaches, is fundamentally constrained by the accuracy of the physical forward model. Forward model mismatch is an inherent real-world challenge arising from discrepancies between the theoretical model and the experimental system, as well as perturbations in physical parameters, leading to cumulative reconstruction errors. In this work, we present a diffusion-model-based reconstruction framework that explicitly mitigates such mismatches. The residual discrepancy between measured and model-predicted signals is treated as a learnable correction term, efficiently estimated via a shallow residual network. This network functions as a data-driven guidance module, dynamically compensating for model inaccuracies and enhancing object recovery. The proposed framework demonstrates robust and high-fidelity reconstruction in lensless holographic imaging and imaging through scattering media.