Sber launches Kandinsky 6 Image: the image generator gets faster and learns to restore photos

Sber has updated its image model and built Kandinsky 6 Image into GigaChat. The new version bets not only on text-to-image generation but also on full-fledged editing: from restoring old photos to precisely correcting individual details and preparing images for further animation.

What has changed

Kandinsky 6 Image is available directly inside GigaChat, and for users this is arguably the main practical change. The model can be used for free and with no limit on the number of tasks, which means this is not a one-off demonstration of capabilities but a service built for regular work.

According to Sber, the new version is twice as fast as the previous one, handles complex prompts better and produces more detailed images. This matters beyond people who like to experiment with prompts. The faster the model responds and the more accurately it holds on to details, the closer it gets to being an everyday tool for designers, marketers, small-shop owners and content creators. In these scenarios, generation as a flashy feature matters less than a predictable result that can go into a post, a product card, a banner or a presentation without lengthy rework.

On the technical side, the model now uses a Mixture of Experts architecture. In this approach, each request activates not the entire neural network but only the fragments it needs. For the user, this shows up not in abstract metrics but in faster processing and fewer errors on complex scenes that require holding many conditions at once: style, composition, characters, clothing, background and small objects in the frame.

New use cases

The main difference between Kandinsky 6 Image and an ordinary image generator is its expanded set of editing tools. The model now covers more tasks that previously had to be handled in several apps or by hand, especially when the user needs not to invent an image from scratch but to carefully improve an existing shot. The emphasis is clearly on everyday scenarios, not just on showcasing technological capabilities in promo videos for general users and teams.

• Restoration of old photographs
• Styling photos into different visual formats
• Neural photoshoots based on user images
• Changing clothing and backgrounds in photos
• Retouching and virtual makeup

Sber separately stresses that local edits have become more precise. This matters for anyone who works not with generation from scratch but with editing a finished picture. In such tasks the user usually needs to change one part of the image without wrecking the rest. If the model really preserves the original scene better and changes individual elements more cleanly, that makes it noticeably more useful for personal and commercial use.

Search and animation

Another notable addition is built-in image search. If the model does not know what a character or object the user asks for looks like, it can find a visual sample on its own. In practice, this should reduce misses in situations where a text description is not enough or where a current reference matters more than an averaged interpretation from the training data. For services of this class, it is a way to cut down on invented details and bring the result closer to what the user expects.

Generated images can be animated immediately through the linked Kandinsky Video model, which already runs in the GigaChat ecosystem. This turns the service from a simple image generator into a broader visual builder: the user first gets a still frame and can then quickly turn it into a short moving clip without switching to another product. For Sber, it is one more step toward making GigaChat seen as a universal AI assistant rather than just a text chatbot.

What this means

Kandinsky 6 Image shows that competition in generative graphics is shifting from the question of who draws best from a prompt to a more applied one: who gives the user a complete set of tools in one window. For the market, this means more mass-market services where generation, retouching, reference search and basic animation come together in a single workflow, without extra friction or switching between different apps, for a mass audience.