Nvidia introduced a neural network with video generation from a text description

The American company Nvidia at the IEEE Conference on Computer Vision and Pattern Recognition introduced a new version of the neural network that generates video from a text description. According to the developers, the training is going very fast even compared to the previous month.

Among the examples Nvidia showed were images for the queries “snowman in a snowstorm”, “dressed fox dancing in the park”, “lone traveler in a foggy forest at dawn”, and others. The video is created in either 512×1024 or 1280×2048 resolution and consists of 113 frames about five seconds long. The neural network takes into account about 4.1 billion parameters, of which 2.7 billion are trained on video. Formerly Nvidia added the function of scaling the video image in browsers.

Evgeny Fedunenko