In the two days following the rollout of Kandinsky 2.2, Sberbank reported that users had generated three million unique images. This surge underscores how quickly the upgraded neural network has captured interest across the user community.
Kandinsky 2.2 promises high-resolution, photorealistic outputs, with notable improvements in portrait realism compared to its predecessor. Users now have the option to adjust the image aspect ratio during rendering, moving beyond the previous square formats. This flexibility enhances the ability to tailor visuals for social media, marketing, and personal projects across North American audiences.
The latest version also introduces new functionality for stickers. Kandinsky 2.2 can craft unique sticker designs that integrate smoothly with Telegram sticker packs. In addition, a Telegram bot powered by Kandinsky 2.2 can replace items or elements within an uploaded image according to a text description, all while preserving the original composition thanks to the ControlNet-based architecture. This makes it possible to experiment with composition while keeping the core layout intact.
Beyond these features, Kandinsky 2.2 can generate illustrations from multiple images, stylize drawings based on textual prompts, and produce outputs that resemble provided references. It supports inpainting to fill missing parts and offers outpainting to extend visuals beyond existing boundaries. The system is described as understanding Russian prompts, enabling image creation in more than 20 distinct styles. These capabilities broaden the range of creative possibilities for users in Canada and the United States who rely on diverse visual styles and language support.
Testing opportunities are provided through multiple channels. Model samples can be explored on platforms like FusionBrain.AI, via a Telegram bot named for artist-style interactions, and on compatible smart devices integrated with social apps. Availability also extends to the Salyut mobile application and a platform called ML Field. This multi-platform access makes Kandinsky 2.2 readily available to a broad audience across North America.
The neural network behind Kandinsky 2.2 was developed and trained by Sber AI researchers in collaboration with scientists from the AIRI Artificial Intelligence Institute. The training leveraged a joint dataset combining resources from Sber AI and SberDevices, reflecting a concerted effort to advance practical AI art generation for wide deployment.
Source: Sberbank press service. The company indicates that the technology can be tested on the Kandinsky 2.2 promotion page models, as well as through Telegram interfaces and on Sber smart devices that run VKontakte integrations. Users are invited to explore the capabilities on supported platforms and observe how the system handles various prompts and styles. The collaboration between in-house researchers and external partners demonstrates a collaborative approach to scaling AI-driven creative tools for mainstream use. The information provided here reflects the latest public updates from Sberbank regarding Kandinsky 2.2 and its ecosystem of integrations and testing avenues.