Flowchart of the proposed method. The lens potential is modeled with a continuous neural field that takes as input any position in the image plane and outputs the value of the lens potential at that position. Alternatively, an analytical profile (e.g., a SIE) models the smooth component of the lens potential, while the neural network captures deviations from that smooth component. The input coordinates are first passed through a Fourier feature mapping (γ_σ) to increase the dynamic range of the recovered features. The different model components follow Eq. (4) except for the blurring operator, omitted to avoid clutter. Since the model is fully differentiable, automatic differentiation is used to compute the exact gradient of the highly nonlinear loss function.