Workflow for the training of DeepHMI in the reverse diffusion process. We send a randomly selected image x_t, together with an HMI LOS magnetogram, to U-Net, where the HMI LOS magnetogram is temporally and spatially aligned with the corresponding GST/NIRIS LOS magnetogram x₀. U-Net predicts the noise ${\widehat{ϵ}}_t$, which is compared with ϵ_t generated by the forward-diffusion process using a loss function to optimize the weights of the neurons in U-Net through backpropagation.