Sber presented the third model of the Kandinsky neural network

The new generative model can create high-quality images in a few seconds according to their textual description in natural language, a Sberbank representative told RB.RU.

You can also mix multiple drawings, change them according to a text description, generate images similar to the given one, fill in the missing parts of the image, and form images in endless canvas mode (paint / paint).

The model recognizes requests in 101 languages ​​(including Russian and English) and can draw in various styles.

Kandinsky 2.1 inherited the weights from the previous version, which was trained on 1 billion text-image pairs and additionally trained on 170 million high-resolution text-image pairs. It also took into account a separately collected dataset of 2 million pairs of images with descriptions in traditionally difficult areas for neural networks: text and people’s faces.

The new version of Kandinsky 2.1 contains 3.3 billion parameters instead of Kandinsky 2.0’s 2 billion and uses not only an encoded text description, but also a special representation of the image using the CLIP model. In this way, the neural network generates an image representation based on textual information and feeds it to the input of the main generative model.

• Sber first introduced a model for generating images based on Kandinsky’s text description in June 2022: it was an improved version of the already created ruDALL-E neural network.
• In November of the same year, Sber presented an improved multilingual Kandinsky 2.0 model, which worked with 2 billion parameters and made it possible to create images in 20 different styles, for example, Renaissance, Classicism, Khokhloma.