Physics-informed diffusion enables high-fidelity multimode fiber imaging under dynamic perturbations

Abstract The convergence of data-driven optics and intelligent imaging has renewed interest in multimode fibers (MMFs) as compact, high-capacity platforms for optical communication and endoscopic imaging. However, the perturbation-dependent transmission of MMFs makes the input-output mapping non-stationary, therefore, reconstruction models that are learned once must be conditioned on up-to-date physical priors to remain accurate. We present a physics-informed diffusion-based network (PID-Net) that enables robust image reconstruction through dynamically perturbed MMFs. By integrating dynamically updated, wavelength-multiplexed transmission-matrix priors, PID-Net embeds the deterministic optical transmission law into the diffusion-based framework, enabling fast recalibration and physically consistent reconstruction under dynamic perturbations. Experiments on meter-scale multimode fibers under dynamic bending demonstrate that PID-Net achieves a fourfold improvement in reconstruction quality, preserving rich structural details where conventional methods completely fail. These results demonstrate that PID-Net markedly improves robustness to fiber perturbations and provides a practical, physically grounded route toward deployable multimode-fiber imaging.