Kandinsky: The Evolution, Capabilities, and Global Impact of Sber’s Generative AI
Named in honor of the pioneering Russian-French abstract painter Wassily Kandinsky, the Kandinsky neural network family has evolved from a regional text-to-image experiment into a sophisticated, multimodal suite of open-source models capable of generating both high-resolution imagery and dynamic video. Developed by Sber AI alongside the Artificial Intelligence Research Institute (AIRI) and SberDevices, Kandinsky stands out as one of the prominent non-Western entries in the global generative AI landscape, frequently compared to models like Midjourney, Stable Diffusion, and OpenAI’s DALL-E.
1. From Humble Beginnings to Multilingual Diffusion
The Kandinsky project kicked off in June 2022, initially relying on an autoregressive architecture similar to Sber’s earlier ruDALL-E. However, the real turning point arrived with Kandinsky 2.0, which pivoted the model family toward latent diffusion principles and introduced robust multilingual capabilities spanning over 100 languages.
Leave a Reply