Sber's Kandinsky now renders Russian text
Sber has updated its Kandinsky image model so it can draw Cyrillic text inside generated images, which addresses a long-standing weakness of image generators for Russian-language users.
Sber's Kandinsky team announced on Habr that its image generation model can now produce Cyrillic lettering. The text is meant to be part of the scene rather than laid over it. The post's examples include words painted on a wall, cut from wood, cast in metal, embroidered, knitted, or spelled out in rose petals.
The team says it first built a dataset of more than 10 million images containing Russian text in a wide range of styles, so the model would learn printed, handwritten and other letterforms. It then trained Kandinsky on that data to generate Cyrillic text natively, without separate add-on modules. A final round of fine-tuning used a smaller expert dataset that designers and artists selected and checked.
The developers say the model is still unreliable for some prompts. Long text, text that mixes Cyrillic and Latin script, and prompts that describe the subject or background in detail may take more than one attempt. They say short prompts work faster and look better, and that naming textures and lighting, such as stone, ice, glass or backlighting, tends to help.
The post gives no benchmarks and no comparison with other models, and it says nothing about model size, architecture or API access.
The update is live in the Kandinsky Telegram bot, in all GigaChat bots (Telegram, VKontakte, Odnoklassniki and Max), and in the web version.