Kandinsky Neural Network turned two years old

Today marks two years since Sber launched the Kandinsky neural network, which has enabled users around the world to generate more than 600 million images. This event marks a significant contribution to the development of technologies for generating images from text descriptions.

History of Kandinsky

Two years ago, on June 14, 2022, Sber unveiled its first neural network for generating images from text in Russian – Kandinsky. This innovation was made possible through the collaborative efforts of researchers and engineers from Sber AI and SberDevices, supported by scientists from AIRI Institute. The model’s name was chosen in honor of Wassily Kandinsky, the famous Russian abstractionist painter and art theorist.

The Evolution and Development of Neural Networks

Kandinsky is a continuation of the ruDALL-E neural network introduced on November 2, 2021. Unlike its predecessor, which was trained on 200 million images with Russian-language descriptions, Kandinsky demonstrates significantly higher generation quality. The model has undergone several phases of updates, resulting in Kandinsky 2.0, 2.1, 2.2, 3.0 and 3.1 versions, as well as the Kandinsky Video family of models for video-to-text generation.

Technical advances and features

Kandinsky utilizes a variety of image generation approaches, including autoregressive transform models and diffusion models with image prior-block. These techniques produce high quality images that are culturally and historically contextualized. For example, Kandinsky can generate images in Gzhel and Khokhloma styles. In addition to image generation, the model has editing and instructional generation capabilities, including additional drawing, correction, blending, stylization, and object replacement.

Popularity and Recognition

Kandinsky 2.1 has become one of the fastest growing neural networks in the world, attracting more than 1 million unique users in just four days after its release on April 4, 2023. Kandinsky 2.2 ranked first in growth rate and was the second most popular among developers after Stable Diffusion according to AI resource Hugging Face. Kandinsky also became one of the most mentioned Russian neural networks in social media in 2023-2024, according to a Brand Analytics study.

And Kandinsky is one of the most mentioned Russian neural networks in social media in 2023-2024, according to a Brand Analytics study.

Real World Applications

The Kandinsky neural network is used in various projects that help preserve and popularize Russia’s cultural heritage. As part of the Revived Collection project, the neural network has helped restore images of lost paintings by famous Russian artists. Murals created with the help of Kandinsky 2.2 have decorated schools in Volgograd, and the Imperial Porcelain Factory has released a collection of mugs with drawings created by the neural network.