Androids may well dream of electric sheep, but it turns out that the dreams of artificial neural networks are a lot less fluffy. Last month Google revealed an interesting by-product of its image recognition research (opens in new tab): what happens when you feed an image into a neural network and then ask it to enhance whatever it detects.
If you do this with the part of the system that's programmed to search for actual objects in an image, then get it to keep on doing it, you end up with a feedback loop that can conjure up unexpected objects from the most unlikely details, resulting in hallucinatory, dream-like artwork. Its initial 'Inceptionism (opens in new tab)' gallery proved so popular that a couple of weeks later Google released an open-source code example so that people could try it out for themselves.
DeepDream (opens in new tab) isn't for the faint-hearted; at the very least you'll need some coding skills and a familiarity with Python to get anything out of it. However there's already a web app that you can use to apply DeepDream effects to your own images.
Dreamscope (opens in new tab) is an easy-to-use app that enables you to upload a picture and enhance it using a variety of DeepDream techniques. There are various psychedelic edge enhancement tricks, but the fun comes when you use the Trippy setting: that's the one that'll turn beautiful landscapes into Hieronymus Bosch-style scenes littered with nightmare mutants.
To really guarantee a sleepless night, we'd recommend saving the output and then feeding it straight back into Dreamscope – perhaps using a different effect – and then repeating the process until your brain can't take the horror any more.
Words: Jim McCauley
Jim McCauley is a writer, editor and occasional podcaster, and is available for space parties.
Like this? Read these: