Dreamix: unprecedented result of experimenting with video generation presented by Google

Google has presented the first diffusion-based method for performing text-based motion and appearance editing of general videos.

Dreamix enables to edit video, generate video following an image or a set of pictures, change an action and more. There are some illustrations published.

Video Editing

Image-to-Video

Subject Driven Video Generation

There are several visible disadvantages. First, the neural network produces a poor-quality picture. Second, animals look unrealistic in generated videos, but inanimate objects are clearly better.

So far, the Dreamix neural network is an internal Google development. It is unclear whether it will be available to the public in the near future, but the beginning looks very intriguing.