Global mean redshift residuals and σ_MAD (defined in the main text) for all the image translation models. The results from the capsule network implementation without using the morphological class labels are also shown, denoted as “CapsuleNet_S2S (no morph.)”. The redshift predictions were made for the generated test images using redshift estimation models trained with either the original or the generated ɀ-training images, denoted as “original-generated” and “generated-generated”, respectively. These results are compared with the predictions for the original SDSS test images by a model trained with the original SDSS ɀ-training images (i.e. the baseline). The error-bars are too small to be visible. The mean residual and σ_MAD from the diffusion model in the original-generated case are dramatically beyond the plotted ranges (i.e. 0.193 and 0.214, respectively) and indicated by the red arrows.