Google has introduced the Lumiere neural network, which can create realistic videos based on text prompts. The new technology uses a spatial-temporal diffusion architecture, which allows you to calculate not individual images, but the entire video. This allows you to create smoother and more natural motion than with traditional video generation methods. This is reported by snob.ru with a link to the project website.

Lumiere can create videos up to five seconds long. She can also convert static images into dynamic ones, copy the style of other videos, and edit existing videos.

The neural network, trained on 30 million videos, is not yet available to the public. However, according to journalists from the VentureBeat portal, the full launch of Lumiere will allow Google to compete with other neural networks that create videos, such as Runway, Pika and Stability AI.