According to the company, the new technology handles portraits better and can create stickers more efficiently than its predecessor.
Kandinsky 2.2 is the latest improved version of the neural network of the same name, which Sber introduced in June 2022. This is not the first update: Kandinsky 2.0 was introduced in November of the same year, and Kandinsky 2.1 in April 2023.
The new version of Kandinsky 2.2 can be tested for free on Sber's website, on the Fusion Brain and ML Space platforms, via bots in Telegram and VKontakte, and with the voice command "Launch Artist" on Sber smart devices and in the Salute app.
The developers have further trained Kandinsky 2.2 on a dataset of 1.5 billion text-to-image pairs, 300 million more pairs than in the Kandinsky 2.1 dataset. Because of this, the neural network can produce more photorealistic, high-resolution images, better handle portraits, and is able to change aspect ratios.
The new version of the Kandinsky model also added ControlNet, a "special controlled change building block". This architecture allows you to change individual elements on already generated images while maintaining the original composition.
In addition, the model has been trained to generate stickers that can be used to create custom collections on Telegram.
Kandinsky understands queries in 101 languages, including Russian and English, and can be used in more than 20 drawing styles. It can also blend multiple images and complete missing elements. The training took place on the Sber AI and SberDevices dataset under the guidance of Sber AI employees and scientists from the AIRI Artificial Intelligence Institute.
Ailib neural network catalog. All information is taken from public sources.
Advertising and Placement: [email protected] or t.me/fozzepe