Sberbank has introduced Kandinsky 2.1 – an updated neural network for generating images on demand. Now it works better with high resolution and complex objects. Sberbank announced a major update to the Kandinsky neural network, which generates images from text descriptions. The neural network can change existing images on demand, combine several images, draw fragments, create images in the same style as suggested by the user, and more. It understands requests in 101 languages.
Examples:
In version 2.1, the neural network received additional training (on 170 million text-image pairs and a separate collection of 2 million images). At the same time, the encoding model was improved, which made it easier for the neural network to handle high-resolution images, as well as faces and other complex objects. The neural network is available on the website and as a Telegram bot, and can also be activated on Sberbank’s smart devices with the command “Launch the artist”.