Dreaming of bathrooms
As part of the Error and Power residency I wanted to set out to have a computer imagine images and conversely have a person to create an image as a computer would.
Images made by computers and an ‘artificial intelligence’ have come along way in the past five years.
DeepDream is a Google program, created by Alexander Mordvintsev, which help popularised the term ‘dreaming’ to refer to images generated by artificial intelligence. Released in 2015, the bizarre images generated by it seemed to made up of fractals, eyes and dog’s noses.
Since then there’s been an increasingly realism to the generated images. NVIDIA’s ThisPersonDoesNotExist.com generates pictures of people that has ‘imagined’.
Similarly, https://thiscatdoesnotexist.com does the same with cats, although maybe not with the same level of success.
Having separated out a fair number of bathroom images from the downloaded pictures from Zoopla, I gave these to an AI program to see what it would concoct, to see what it bathrooms it would dream of.
The first images the program produced have a kind of painterly look to them, looking like a hazy dream where fixtures and fittings are almost visible.
However, due to the power of my PC and the age of the program I was using, the images are very small – only 256 x 192 pixels. By using a further AI program it is possible to enlarge them. The program imagines what is missing from the image to produce an almost classic police show ‘enhance!’ moment.
This enlarging process is also trained on using a large batch of images. In the case of the program I’m using here, it comes with some fairly small ‘pre-trained’ image sets which probably have never seen a bathroom, not least one with ceiling sinks, adding to the potential for the images it generates to have inaccuracies, glitches. This adds a further layer of dreaminess to the image and both introduces and enhances flaws and glitches in the original image.