Architecture of the CNN. Noisy input images are embedded into a latent space vector by performing a PCA on the wavelet scattering fields s₂^{j₁, l₁, j₂, l₂} with J = 4, L = 8, which is then propagated through one fully connected layer and five convolutional layers, each with a 5 × 5 kernel, batch normalisation and ReLU. The final output of the generative model is produced by a tanh activation function.