Mars Dreams offers a visualization of celestial structures through the mind of a machine. Each iteration of Mars Dreams is a unique latent walk through the invisible connections between various projections of celestial realities that artificial intelligence reveals. Using machine learning and a generative neural network trained on real pictures of Mars, the work imagines the possibility of a universe that would provide further texture to the perception of our own. Data paintings in Mars Dreams series represent past space explorations with machine-generated dreams based on a dataset of satellite imagery from the Mars Reconnaissance Mission provided by NASA. The dataset is then used to teach a computer to understand and generate hallucinatory landscapes of Mars. Over 800 thousand images train a generative adversarial network (GAN) how to dream and imagine the topology of this enticing planet, adding an artistic layer to the recent developments and discussions about imagining life on Mars.