NVIDIA has created a neural network for video generation. She’s like DALL-E 2, only for commercials

A few years ago, when artificial intelligence was only taught to generate texts or perform another routine task, many users believed that creative professions would not suffer from this, since a computer would never be able to draw a picture or something like that. At the moment, the situation has changed dramatically – DALLE-2 or Imagen can generate very realistic images of static content, which is not inferior to the paintings of great artists. Moreover, now it’s time to develop further – NVIDIA has created a model that is able to generate full-fledged videos, and not just static images.

Animation available on click

The latest model from NVIDIA is based on the StyleGAN architecture (generative adversarial network introduced by NVIDIA in 2018), but with many additional components. As a result, it allows you to generate video – the content quite adequately demonstrates moving objects in the frame, as well as changes in the “camera” point of view and various new objects and conditions that arise in the video over time. However, which was to be expected, the technology is far from ideal and works with jerks – even on the videos that were selected by the developers to demonstrate the new model, you can notice obvious deformations of the background and objects when changing perspectives.

Animation available on click

It is also worth noting that all videos are presented in a rather low resolution (128×128), so now it’s too early to talk about their full use. On the other hand, the authors of the model stated that the underlying StyleGAN-V network simply repeats the same content, while the new model is able to generate scenery and objects that appear in the frame over time. At the same time, the system manages to maintain the temporal consistency of the frame, which, theoretically, in the future will allow you to shoot entire films only with the help of AI.

Source: Trash Box

You may also like