A team of researchers at Rutgers University created a machine learning algorithm for churning out hot and ready multi-topping fake pizzas. They called it the Multi-ingredient Pizza Generator, or MPG.
The MPG is a generative neural network, or GAN, using Nvidia's StyleGAN2 architecture—the same machine learning structure that brought us generated furries and plenty of other things that don't exist. Fake pizza generation, however, works as a useful analogy when you're trying to work on AI image generation problems like multiple labels—each ingredient gets a label, and the algorithm can try to assemble the desired pie-output like a recipe.
The researchers published their work on the arXiv preprint server this month. According to the study, they created their own dataset, called Pizza10, based on a pizza-picture dataset developed by a team of MIT researchers and released last year as part of their pizzaGAN project. The MIT team downloaded half a million pizza images by scraping pizza-related hashtags on Instagram, and narrowed it down to 9,213 pizza images with 13 different topping options. They then had Amazon Mechanical Turk workers label the toppings.
Fangda Han, one of the study’s co-authors, demonstrates tweaking style noise controls and adding ingredients in a video:
Maybe they can hook the MPG up to that bad pizza making machine from this year's CES and truly create a monster. Or Domino's can use this data to take a smarter crack at its "send us a shitty pizza picture" promotion and not get owned by &pizza this time.