Denoising of a 100 × 100 arcmin² simulated convergence map (a), which was corrupted with a noise level of σ = 0.15 per pixel (b). The Gaussian denoiser (c), computed with Eqs. (18) and (12), retrieves large scales of the convergence map. Adding the DAE residuals to the Gaussian denoiser (d) allows us to recover both large and small scales of the κ map. This illustrates the Gaussian (analytic) and non-Gaussian (learned) parts of our hybrid score function Eq. (19).