Photometric vs. ground-truth redshift for deterministic and probabilistic models trained and evaluated on different datasets. The top row shows the models evaluated on a sample similar to the training dataset. The bottom row shows the models evaluated on a sample different from the training dataset. NN-1 and BNN-1 were trained on 200,000 galaxies from GalaxiesML. NN-2 and BNN-2 were trained on 81,000 galaxies from TransferZ. The solid black line is a one-to-one line and the dashed black lines correspond to outliers with z_pred > ± 0.15(1 + z_true). The color scale indicates the density of the data points (log scale). All models perform well on the ground truth they are trained on, but show greater scatter and more outliers on a sample different from their training set.