Yandex introduced the YandexART diffusion neural network for creating images and animations

Yandex AI Rendering technology generates images and animations based on text queries. The company says the service “understands the Russian cultural code” and learns about important symbols, characters and personalities from Russian history.

The service creates images and animations using the cascade diffusion method, the company explains. First, the neural network generates images and frames at the user’s request and then gradually increases their resolution, filling them with details.

Deep details and realism are YandexART’s advantages over the previous generation of the diffusion model, which was implemented in Masterpiece, the company says.

To improve the quality of the neural network, the developers increased the size of the training data set by 1.5 times: up to 330 million images with text descriptions. And thanks to the new algorithm, the service better understands the user’s request.

Additionally, to improve the quality of work, additional reinforcement learning is used. Experts manually review the content generated by the neural network and highlight successful and unsuccessful options.

The technology has already been implemented in the company’s services, for example in Masterpiece. And in Yandex.Business, a neural network helps illustrate ads. In the future, YandexART will appear on Yandex.Keyboard.

Previously, Sber taught its Kandinsky neural network to generate animated videos based on a text request. Kandinsky 2.2. Create four-second videos with a resolution of 640×640.