We adopt a data-driven approach to model the longitudinal phase-space diagnostic beamline at the European XFEL photoinjector. A deep convolutional neural network (decoder) is used to build a 2D distribution from a small feature space learned by another neural network (encoder). We demonstrate that the autoencoder model trained on experimental data can make fast and very high-quality predictions of megapixel images for the longitudinal phase-space measurement. The prediction significantly outperforms existing models. We also show that the knowledge learned by the model can be transfered to speed up training of another model with a different injector setup. This opens the door to a new way of precisely modeling a photoinjector and can be possibly extended to the whole accelerator.