PixelQuiz, a video game that allows players to guess an image by looking at its pixelated version, was launched in February this year. Developed by Robyn Dubuc, the images in the quiz are designed by machine learning. Each image is unlocked when the user clicks on the centre of the image and is promptly drawn by AI. The video game is divided into multiple categories like movies, video games and famous personalities. The app is still in its early stages. 

Dubuc has previously served as the lead art director at an online video game design studio, XGen Studios. He also said that while the iOS version for the game was freely available, the android version is still in development. There are 657 games in just the video games category. 

Dubuc explains that the model was trained using a massive set of images pulled from the internet on varied subjects. The image starts as complete noise and then gradually alters itself after many iterations. The model uses self-supervised learning to then train itself so that the image matches itself to the prompt in its training set. Post this; it assesses how well it did the job and scores itself. The next time the image makes random adjustments to get an improved score. 

via Gfycat

Dubuc himself commented on the game, expressing amazement that the images were created entirely by machine learning and there was no pixel artist involved. He added, “I would explain more about how it worked, but honestly I don’t understand it myself.”