Sberbank presented Russia’s first neural network for creating video – Kommersant

2023-11-22 13:11:51

Sberbank (MOEX: SBER) presented Russia’s first neural network for creating videos using text descriptions – Kandinsky Video. It can generate a video sequence up to eight seconds long at a frequency of 30 frames per second, said Alexander Vedyakhin, deputy chairman of the bank’s board of directors, during the AI ​​Journey international conference.

“We recently trained Kandinsky to create animated videos using text descriptions, and today we are introducing a model of a completely different level – the first model in Russia for generating full-fledged videos from text. This is an important contribution to the development of Russian generative neural networks,” said Mr. Vedyakhin (quoted from press release Sberbank).

Kandinsky Video consists of two blocks. One of them is responsible for creating key frames that make up the structure of the video plot, the second is responsible for generating interpolation frames that ensure smooth movements in the video.

The neural network creates videos with a resolution of 512512 pixels and different aspect ratios. The model was trained on a dataset of more than 300 thousand text-video pairs. It will take up to three minutes to generate the video.