CNN metrics as a function of redshift label and i-band magnitude for the u^* grizy and ugrizy validation sets, and for the u dataset inferred from the u^*-trained model. Using either set of bands for training makes no difference but inferring one from the other significantly degrades the performance. In this and all similar figures, the redshift and magnitude distributions are overlaid in gray in the corresponding upper panels, and the 1 σ error bars are based on 100 bootstrap samples.