Built by — Philip Pries Henningsen
Philip is a machine learning engineer who enjoys working with natural data and problems. By that, I mean problems that are directly relatable to humans — such as vision, audio, language etc. For the last four years, I have mainly worked with computer vision problems, which is anything that involves understanding images or video.
Neverending Catalogue is a technical prototype which would create computer-generated bedrooms to act as an inspiration database, essentially taking on the role of an interior designer for you to consult at your leisure.
00:00 / 00:00
Latent space walk from the final stage of the self-trained model
It involves using generative adversarial networks (GANs) to generate images of completely new bedrooms using pre-existing photographs of rooms lifted from the IKEA catalogue. This experiment tried to answer the questions: Is there a particular "IKEA style" for decorating a room? And can a machine learning model understand it and generate new rooms?
00:00 / 00:00
Latent space walk from 3rd training round
Sample of one of the best results from the 2nd round RunwayML trained model
Sample of one of the best results from the 1st round RunwayML trained model
Images of bedrooms are trained on top of a generative adversarial network (GAN) which has already been exposed to thousands of generic bedroom images. The result is that the images created by the network are flawed, but beautifully so. This is exciting. An application of a model like this could be used by those looking to re-decorate, or even to give interior decorators or designers new ideas on what to do with a room and furniture. Even with obvious errors, the wild, unpredictable scenes the algorithm is capable of dreaming up could still be used as inspiration.
By navigating this website you agree to the storing of cookies as described in our terms of use.
View Experiment
View Experiment
View Experiment
00:00 / 00:00