Within a single day of releasing the new version of the Sber Kandinsky 2.2 neural network, the bank’s press service reported that users had generated one million unique images. The most popular subjects were cats, the sea, and anime scenes, illustrating a wide range of interests among the community.
Kandinsky 2.2 enables high-resolution photorealistic image creation, with notable improvements in portrait rendering compared to the previous version. Unlike earlier releases that produced square outputs, this update allows users to adjust the aspect ratio during rendering, giving artists and creators more flexibility. The neural network also demonstrates how to craft distinctive Telegram stickers by blending sticker packs, showing that it has learned to adapt to practical, everyday needs as well as creative experimentation.
Additionally, the Kandinsky 2.2 bot on Telegram can rapidly replace specific objects or elements in uploaded images based on textual descriptions, thanks to its ControlNet architecture. Importantly, it preserves the composition and overall balance of the original image while making targeted changes.
The new Kandinsky 2.2 release can also generate illustrations from multiple source images, assemble a drawing guided by text, imitate a given style, fill in missing parts of an image (internal painting), and even operate in an infinite canvas mode for expansive creations. These capabilities expand the ways users can translate ideas into visual form without needing multiple tools.
According to Sber, the network understands requests in Russian, setting it apart from many overseas counterparts. It supports creation across more than 20 distinct artistic styles, offering a broad palette for creators working in different genres and moods.
Users can test Kandinsky 2.2 on several platforms. Models for the promotion page are accessible via the FusionBrain.AI platform, along with Telegram and VKontakte bots. The system can also be activated using the command Start the Artist on Sber smart devices, within the Salyut mobile app, and on the ML Space platform, showing a multi-channel approach to access and experimentation.
The neural network was developed and trained by Sber AI researchers with support from scientists at AIRI Artificial Intelligence Institute, using a combined dataset built from Sber AI and SberDevices resources. This collaboration highlights the emphasis on robust data foundations and practical training pipelines, aiming to deliver reliable and versatile creative tools for a broad audience. Attributions for the development and testing processes reflect ongoing cooperation between industry and academic partners to push forward the capabilities of generative models in a responsible and user-friendly manner.