ABSTRACT Machine learning (ML) is one of two standard approaches (together with SED fitting) for estimating the redshifts of galaxies when only photometric information is available. ML photo-z solutions have traditionally ignored the morphological information available in galaxy images or partly included it in the form of hand-crafted features, with mixed results. We train a morphology-aware photometric redshift machine using modern deep learning tools. It uses a custom architecture that jointly trains on galaxy fluxes, colours, and images. Galaxy-integrated quantities are fed to a Multi-Layer Perceptron (MLP) branch, while images are fed to a convolutional (convnet) branch that can learn relevant morphological features. This split MLP-convnet architecture, which aims to disentangle strong photometric features from comparatively weak morphological ones, proves important for strong performance: a regular convnet-only architecture, while exposed to all available photometric information in images, delivers comparatively poor performance. We present a cross-validated MLP-convnet model trained on 130 000 SDSS-DR12 (Sloan Digital Sky Survey – Data Release 12) galaxies that outperforms a hyperoptimized Gradient Boosting solution (hyperopt+XGBoost), as well as the equivalent MLP-only architecture, on the redshift bias metric. The fourfold cross-validated MLP-convnet model achieves a bias δz/(1 + z) = −0.70 ± 1 × 10−3, approaching the performance of a reference ANNZ2 ensemble of 100 distinct models trained on a comparable data set. The relative performance of the morphology-aware and morphology-blind models indicates that galaxy morphology does improve ML-based photometric redshift estimation.