How to use the Kandinsky neural network

What is Kandinsky

Kandinsky is Sber's neural network for generating images. It is capable of creating images based on text queries, as well as presenting variations of ready-made images and combining styles from different frames.

The system identifies requests in different languages, including working reliably in Russian. The latest version of the service at the moment is Kandinsky 2.2. The neural network takes into account additional parameters when generating the result, including background and style.

What the Kandinsky 2.2 neural network can do

As mentioned above, Kandinsky can not only provide images on demand, but also create images by mixing different concepts or styles. The Sber neural network supports several operating modes. In standard it generates the result by introduced text message. When combining frames, it analyzes the two and composes a new one from them.

You can also “feed” one finished painting or photograph to the system by adding the necessary characteristics. In this case, Kandinsky will create a new image, given the visual example and the promt at the same time.

In addition, the service supports outpainting mode, or finishing drawing. This function allows you to supplement the finished frame with new details that were not there before. Another mode of operation is style transfer. With its help, you can use some of the details of the original image in the generated image.

How to use Kandinsky 2.2

The neural network is available through several services in different formats. So, Kandinsky can be tested on the website Fusion Brain. There you can generate pictures using text commands, as well as use the finishing tool.

Also available on Telegram official bot Kandinsky. With its help, you can create images based on text, mix two different pictures, transfer the style and create variations of the finished frames.

In addition, on ruDALL‑E website There is a form for creating images with basic settings. The Kandinsky neural network is also integrated into the voice assistant “Firework» from Sber. Here you need to run the “Turn on Artist” skill to generate pictures. In addition, the service can be used via bot "VKontakte" and on official website "Sbera".

The Telegram bot just needs to specify the operating mode with the corresponding button, and then enter the request text or upload the necessary images. The service is free and provides results quite quickly. Failures and errors occur rarely - with a very large number of simultaneous commands from users.

The tool for editing and expanding frames in Kandinsky is only available on the website Fusion Brain. In addition, there is a wide area with an image for work, a text field for promta and a drop-down menu with dozens of styles. From the list you can choose one of the popular examples - from cyberpunk to Soviet cartoons.

The style does not have to be marked in the settings; it can be specified in the text request. You can even write an option that is not yet in the basic list. In this case, you should leave the “No style” option in the menu.

For pictures, you can choose one of the available aspect ratios and resolutions. The neural network produces frames with dimensions of 1,152 × 768 pixels, 1,024 × 1,024 pixels, 680 × 1,024 pixels and vice versa, 576 × 1,024 pixels and vice versa.

Finishing allows you to form pictures from small ideas. It is enough to select part of the finished frame and an empty area, and then enter a text command that Kandinsky will determine exactly how to expand the specified frame, adding details or continuations to it objects.

When working on graphics projects, you can quickly generate new ideas using neural networks "Sbera". To do this, the Eraser tool, or Erase, in Fusion Brain is useful. It is enough to erase part of the finished frame, and then add new elements to the free space according to the text description. In this case, it will be possible to maintain the picture in the same style or combine different concepts.

When transferring a style from an existing frame to a new one, Kandinsky allows you to use the poses of people from a photograph or painting, as well as the general outlines of the original image. For example, in a portrait it will be possible to replace one person with another, while maintaining the overall composition and background. This algorithm works through a ⁠bot in Telegram.

When mixing two images, the system does not preserve the construction or arrangement of objects. The merger occurs randomly, which sometimes leads to unexpected results and new ideas.

What are the disadvantages of Kandinsky 2.2

Kandinsky 2.2 does a much better job of generating realistic frames than previous versions. The results are similar to a popular service Midjourney, but are still inferior in level of detail.

To get good results, you need to experiment with the parameters and descriptions of queries. From time to time Kandinsky produces images with errors. For example, the system displays architectural objects familiar to many in a strange way. But this usually happens if you introduce too long industrial parts with a large number of minor details.