The model reconstructs the original data from the encoding (and the noisy input). This time, it uses a different loss function.
-