Figure 2.

Image transformation network incorporated with the residual dropout mechanism in the training mode and the inference mode. The RD-activation code illustrates the reconfiguring of the residual dropout at the inference mode to amplify the latent space stochasticity without any model retraining.