The new version of Sber’s generative model – Kandinsky 2.2 – can generate photorealistic images in higher resolutions and change the aspect ratio. Portraits now have significantly better quality. To update the neural network, Sber used a dataset of 1.5 bn text-image pairs, which is 300 mn more than had been used to train the previous version – Kandinsky 2.1.
The new version of the model can generate stickers to assemble fully-fledged sticker packs for Telegram. Kandinsky 2.2 bot can now change individual objects or elements on the image based on text descriptions while preserving the composition of the original image, thanks to ControlNet, an embedded dedicated managed change structural block.
Kandinsky 2.2 understands prompts in Russian and English, can paint in more than 20 different styles, blend several images, apply styles, generate images similar to input images, and support inpainting/outpainting. Alexander Vedyakhin, first deputy chairman of the Executive Board, Sberbank, “We continue to improve our neural network to create unique content. The new version provides more options to users and generates images of higher quality. Not only does our neural network try to imitate human creativity, but it can now create new artistic meanings and interpretations. As was the previous version, Kandinsky 2.2 is publicly available, so you can test it for free.”
You can try Kandinsky 2.2 in a Telegram bot, on the model’s promo page, at fusionbrain.ai, on ML Space in DataHub, a pre-trained model and dataset hub, in the Salute mobile app, and on Sber smart devices with the “Launch the Artist” voice command. Kandinsky 2.2 is a neural network developed and trained by the Sber AI team, which partnered up with the Artificial Intelligence Research Institute (AIRI), using a consolidated Sber AI and SberDevices dataset. The previous version, Kandinsky 2.1, had become the world’s fastest-growing neural network with one million active users engaged four days after launch.