Our text-to-dream pipeline takes combinations of words and uses them to generate inputs for the artificial intelligence. The algorithm runs iteratively on its own outputs and applies some manipulation after each iteration, giving us an endless stream of impressions that explores the set of things the network knows.

If we start this process from an image with random noise, the result becomes a pure result of the neural network... and at the end a lot of images come out, which you don't even recognize anymore, similar to the confusing things, which we sometimes really dream.

That is what we describe with "dreaming".