Wholly differentiable virtual lens for ultrathin-plate broadband achromatic imaging

Achromatic imaging seeks to capture the colorful world without chromatic aberration. However, traditional broadband achromatic imaging faces challenges such as complex design, high manufacturing cost, and large volume, due to the gap between physical systems and computational models. Here, we reformulate the concept of a virtual lens as a learnable regularization module within a fully differentiable and physics-encoded computational imaging framework for broadband achromatic reconstruction. Instead of relying on intricate optical design, our method constrains the solution space with wave-propagation consistency priors. A spectral complex-wave estimator lifts RGB measurements into a multi-wavelength amplitude and phase latent field, while a structured regularizer trained end-to-end with the reconstruction network enforces consistency with acquisition physics and suppresses chromatic artifacts. Validations on refractive and diffractive singlet systems across wide fields of view and varied illumination demonstrate up to 14.8 dB peak signal-to-noise ratio (PSNR) improvement and strong generalization to in-the-wild scenes.