Defocus deblurring is typically trained on real sharp–blur pairs, but focus breathing and camera shifts during capture misalign them, leading to over-smoothed restorations. We present a misalignment-free framework that synthesizes blur directly on the sharp image, so training pairs remain aligned by construction. An optically guided implicit neural blur engine predicts depth-dependent kernels, and adversarial learning adapts only this engine toward real defocus.

Seunghoon Lee earned a B.F.A. in Photography (2024) and a Master's degree in Imaging Science (2026). He is currently a Ph.D. student in AI Imaging within the Department of Imaging Science and Arts at the Graduate School of Advanced Imaging Science, Multimedia & Film, Chung-Ang University. His research interests focus on low-level and physics-based vision and generative models.