According to the company’s press service, Yandex introduced Yandex Artificial Intelligence Rendering Technology (YandexART), a diffusion neural network that creates images and animations based on user text requests. It will help with business tasks and creativity.
YandexART creates images and animations using the gradual propagation method. As the company explains, the neural network first produces images and frames as per the user’s request, then increases their resolution and adds details in several stages. At the same time, YandexART creates more realistic and detailed graphics compared to the previous generation of the diffusion model used in Masterpiece. According to the results of internal evaluation by Yandex raters, in 60% of cases the results are better.
“He understands the Russian cultural code; He knows the famous places and cities of the country, outstanding personalities of different eras and familiar from childhood cartoon and fairy-tale characters. For example, Cheburashka, heroes and Babu Yaga. The neural network is already integrated into the Masterpiece application and helps display ads in Yandex Business, and will soon appear in Keyboard and other company services,” the message states.
It is stated that to improve the quality of the neural network, the developers increased the size of the training dataset by 1.5 times, that is, up to 330 million images with text descriptions. At the same time, aesthetic images selected using three filtering modes are used for YandexART training. The neural network also examines features of faces, eyes and hands.
Additional reinforcement training is used to improve the quality of production. Thus, the company’s experts simultaneously look at several hundred or thousands of images created using YandexART and note the most successful and unsuccessful among them. This allows the neural network to better understand how to render images better and more accurately.