An enhanced text-to-image model built into Kandinsky 3.1 by Sber is now accessible to all users without restrictions. At the international industry fair Innoprom. Central Asia, Alexander Vedyakhin, First Deputy Chairman of the Board of Directors of Sberbank, described this expansion as a significant milestone.
The updated Sber neural network version has been trained on a dataset of aesthetic images, enabling sharper rendering and making outputs more realistic, creative, and vibrant. This improvement aligns with feedback from early evaluators who tested the system in its improved form earlier in the spring.
Vedyakhin noted that designers, artists, and bloggers were among the first to assess Kandinsky 3.1’s enhanced capabilities, reporting strong impressions about its potential for creative work.
Positive responses and constructive comments from these early users contributed to refinements in the neural network, with ongoing iterations aimed at expanding practical usefulness for professionals in the creative sector.
As of today, Kandinsky 3.1 is available to all users. The model is positioned as a tool that unlocks creative potential and supports professionals in achieving more efficient development within the creative industry, according to Sberbank.
The release also introduces a streamlined visual creation process through a query-optimization feature, described as beautification, which simplifies prompts and speeds up results.
User expertise in prompt engineering is no longer a prerequisite. By entering a few descriptive words, the built-in GigaChat Pro language model enhances and enriches the prompt with details, allowing the system to produce refined imagery without requiring advanced technical input from the user.
The training approach and the high-quality dataset have contributed to improvements in the inpainting capability, enabling editing of specific parts of an image with greater precision.
The system now includes a fast Kandinsky 3.1 Flash model that can be used directly in the main Telegram bot, reducing rendering time by more than tenfold in this release.
Capabilities can be explored on the fusionbrain.ai platform, and through Telegram and VK bots, as well as in the Sberbank Online mobile application. Additional updates are planned to broaden the range of outputs and workflows for users across different regions and industries, with future releases anticipated to expand 4K image creation and video generation based on Kandinsky Video 1.1 text prompts [Sberbank press materials].