Sunday, August 7, 2022
HomeTech and GadgetsThe Kandinsky neural network creates images from a text description in Russian...

The Kandinsky neural network creates images from a text description in Russian – News

The SberDevices team presented Kandinsky, a neural network model for generating images from a text description in Russian. It is capable of creating a wide variety of content types, from promotional materials and illustrations to architectural design and digital artwork.

The neural network ruDALL-E became the basis of the Kandinsky algorithm . Sber specialists further trained it with the help of 179 million images with text descriptions. The algorithm, according to its creators, is able to generate content with any aspect ratio, accurately conveying the features of various types of textures, as well as additional effects such as shadows and reflections.

The process of creating the desired image is carried out in three stages. First, Kandinsky generates a given number of images, then the ruCLIP Large model selects the most “successful” of them that best match the text description. Finally, the third algorithm scales the finished result without losing detail.

You can already test Kandinsky in action in the Salyut mobile app, as well as on SberDevices smart devices using the “Turn on the artist” voice command.

Html code here! Replace this with any non empty raw html code and that's it.


Most Popular