Kandinsky: Sber's AI Image Generator for Pictures and Video in 2026

Most of the AI image tools you hear about come from the US: Midjourney, ChatGPT, Google's Nano Banana. Kandinsky is one of the biggest you've probably never tried. Sber, the Russian bank behind it, has been building it since 2022, and in 2026 it got two serious upgrades: a faster image model that edits photos precisely, and a video model that adds sound on its own.

I'll be upfront. Kandinsky is built first for a Russian-speaking audience, and the easiest way in, the GigaChat assistant, may not let you in from abroad. I'll flag every spot where that matters. Two things make this guide useful anyway. First, Sber publishes the Kandinsky models openly, so you can run them on your own machine or through a hosting provider. Second, the prompt formula, the 24 prompts, the editing techniques, and the checklist work the same in any image generator you already use.

Section 01What Kandinsky is in 2026

Kandinsky is a family of Sber's AI models that turn text into an image or a video. It's named after the painter Wassily Kandinsky. The first version came out in 2022 and drew the way early AI did: six fingers, melting faces, mush instead of a background. In four years it grew into a tool you can comfortably use for post covers and ad visuals.

If Sber is new to you: it's Russia's largest bank, and over the last few years it has built a whole AI stack around it. GigaChat is its chat assistant, roughly what ChatGPT is for OpenAI, and Kandinsky is the artist living inside it.

Three things matter to you as a user. Kandinsky understands Russian natively, so Russian-speaking users write exactly the way they think. It knows everyday Russian life: an apartment block courtyard, a country dacha, a samovar, a school classroom, and it doesn't confuse them with American equivalents. And it's built right into GigaChat, so Russian users don't need a separate site, an app, or a payment method.

A quick map of versions, so you don't get lost in search results that still talk about Kandinsky 2.2 and 3.0:

In GigaChat you don't pick a version. It always runs the latest one: open the chat today and Kandinsky 6.0 Image draws your picture.

One more detail sets Kandinsky apart from many Western competitors. The new model uses a "mixture of experts" design: each request wakes up only the part of the network best suited to it. That's where the speed comes from. An image usually arrives in 20-40 seconds, while free tiers of other services can keep you waiting for minutes at peak hours.

Now what Kandinsky does worse. Its photorealistic people are good, but in my view still short of Google's Nano Banana: skin sometimes comes out too smooth, as if someone applied a heavy filter. Long Russian lettering on images comes out with mistakes. It's reluctant to imitate a specific living artist. For covers, illustrations, and product shots, none of this gets in the way very often.

Section 02Where to use Kandinsky: 5 ways

Sber no longer promotes a separate "Kandinsky website" for regular people: every new feature shows up in GigaChat first. So the right answer to "where is Kandinsky?" is almost always "wherever GigaChat is." Here's every entry point.

If GigaChat lets you in, start with the giga.chat website on a computer. It's the easiest place to upload photos, compare versions, and download the result at full quality. Keep the Telegram bot on your phone for quick jobs, like "draw a cover for a post about burnout" while you wait for your coffee.

About Fusion Brain. Many older guides send you to fusionbrain.ai, the platform where Kandinsky models used to live and where developers connect to them. For everyday image work you don't need it: GigaChat has the same features and more, without a separate registration.

About Telegram bots with "Kandinsky" in the name. Telegram search turns up dozens of them, and most are unofficial middlemen that show ads, ask you to subscribe to channels, or sell a "premium" plan. The official route to Kandinsky in Telegram is GigaChat's bot, @gigachat_bot. Check that the bot carries the official account mark, and never enter card details into a third-party bot.

Section 03Running the open models yourself

This is the part that makes Kandinsky interesting outside Russia. Sber releases its Kandinsky models openly: starting with the Kandinsky 5.0 family in November 2025, the weights and code are published on Hugging Face and GitHub under the MIT license, a permissive license that allows commercial use. Before you download anything, open the model card of the exact version you want and read its license and hardware notes, since each release ships with its own.

Three ways to run them, from easiest to most technical:

• A hosting provider. Some cloud GPU and model-hosting services add popular open models to their catalogs. If you find Kandinsky there, you pay per image or per minute of GPU time and skip the setup entirely.
• ComfyUI. A free visual tool where you connect blocks like "load model," "prompt," and "save image" into a pipeline. It's the usual home for open image and video models, and it's the friendliest option if you have a computer with a strong graphics card.
• Diffusers. Hugging Face's Python library for image and video models. This is the developer route: a few lines of code, a GPU, and you can plug Kandinsky into your own bot or website.

Be realistic about hardware. Image and especially video models need a graphics card with plenty of video memory. A regular office laptop won't do it. If you don't have a gaming-class or workstation GPU, a hosting provider or a rented cloud GPU is the practical path.

Is it worth the effort? If you just need a cover for Tuesday's post, no: Nano Banana or ChatGPT will get you there faster. The open route makes sense in three cases: you want to generate images in bulk without per-image fees, you want to build image generation into your own product, or you want full control over where your data goes. For everyone else, read the rest of this guide for the prompts and techniques, which work in any generator.

Section 04Is it free, and who owns the images

For regular users, images in GigaChat are free. With the release of Kandinsky 6.0 Image, Sber stated plainly that there's no limit on the number of images you create. In practice you can cycle through twenty versions of a cover until one lands, and pay nothing. For comparison, most Western services give you a few free images a day, then ask for a subscription.

What can be limited:

• Video. Clips cost more to produce, so daily limits apply. Sber changes the exact numbers, so go by what GigaChat shows you.
• Speed at peak hours. On weekday evenings an image sometimes takes longer than usual. That's just a queue.
• Filters. Kandinsky won't draw explicit content, violence, politicians, or celebrities. Sometimes the filter fires on a harmless request; more on that in Section 14.
• The developer API. If you connect Kandinsky to your own bot or site through Sber's API, it has its own pricing. That's a separate story from the regular chat.

Now the rights. This is the question I get more often than "how do I draw something": can I put a Kandinsky image in an ad, on a course cover, in a product listing? Short answer: people widely use these images for their own content and to advertise their own services. But the terms are set by GigaChat's user agreement, and Sber updates it from time to time. Before you launch a big paid campaign, open the current version on the site and read the section about generated content. Five minutes of reading is cheaper than a dispute. If you run the open models yourself, the model's license applies instead. For the bigger picture, see my guide on who owns AI-generated content.

Three rules that hold under any agreement:

• Don't depict real people without their consent, especially in ads.
• Don't ask for a copy of someone else's logo, character, or recognizable photo.
• If the image has your face or a client's face, use a photo you have the rights to.

More tools that genuinely work without a subscription are in my roundup of free AI tools.

Section 05Your first image in 2 minutes

Let's make a cover for a post. Almost everyone who runs a channel needs this: a pretty, atmospheric picture with no faces that you can lay a headline over.

Why it worked. The request is short, but it has everything an artist needs: objects, place, time of day, light, colors, style, and format. The line about empty space on the right solves the biggest pain with covers: without it, Kandinsky fills the whole frame and your headline has nowhere to go.

The classic beginner mistake is a one-word request: "draw a therapist." The AI guesses everything else on its own, and you get a stock picture you've seen a thousand times. The more specifics you give, the less Kandinsky has to make up for you.

Section 06The prompt formula for Kandinsky

A prompt is the text of your request to the AI. For Kandinsky, a six-part formula works well. You don't have to fill every part, but every part you skip, the AI fills in on its own, and not necessarily the way you wanted.

A few observations from practice that save time:

• Language. Russian is the model's native language, and Russian-speaking users write prompts in Russian without translating. The open models also take English prompts. If an English prompt in GigaChat gives you something odd, run it through a translator into Russian and compare the results.
• One main idea per frame. "A therapist in an office, a client next to her, a chart on the wall, the sea outside, a laptop with a slide deck on the desk" turns into a pile-up. Pick the main thing and drop the rest.
• Name the style with words a photographer or illustrator knows. "Shot on film," "watercolor on paper," "flat vector illustration," "3D like an animated movie," "minimalism, lots of air."
• State the format explicitly. GigaChat can pick proportions on its own: vertical for a portrait, horizontal for a landscape. For a cover or a story, say it outright: 16:9 or 9:16.
• You can say what shouldn't be there. "No people," "no text," "no logos." Kandinsky understands that.

If you want to go deeper on prompts, my AI prompt library has 100 ready commands for different tasks, and the text approaches there carry over well to images.

Section 0724 ready prompts for real tasks

Copy them and swap the details for your niche. Each prompt starts with "Draw" so GigaChat hands the request straight to Kandinsky. They work just as well in any other image generator.

Look at the last prompt. A samovar, jam in little glass dishes, a dacha veranda: this is exactly where Western AI models stumble and draw something halfway between England and Mexico. Kandinsky makes it recognizable. The lesson carries over to any audience: local details give a picture warmth that stock photos don't. If your clients are in Texas or in Lisbon, name the things they'd recognize, and test which generator draws them most convincingly.

Another 30 ideas for photos, including portraits and AI photo shoots, are in my collection of AI photo prompt ideas.

Section 08Editing photos: point edits

The biggest change in Kandinsky 6.0 is editing finished photos. Earlier versions redrew the whole frame for any request: you asked for a different wall color and got a different person. Now Kandinsky changes only what you asked for. Ask for a new hairstyle or a pair of glasses, and the rest of the frame stays as it was.

What people most often do with photos in Kandinsky:

• Clean up the background: cables, litter, someone's bag on a chair, random people.
• Swap the background: "put me against a bright studio backdrop" or "bookshelves behind me instead of a bare wall."
• Change clothes for a cover: "put me in a navy blazer, don't change my face."
• Extend the frame: "extend the photo on both sides to make it 16:9," when you have a vertical photo and need a horizontal cover.
• Light retouching: even out the light, remove glare from glasses, make the photo a bit warmer.

Section 09One image from three photos, style, and restoration

Three Kandinsky 6.0 features that few people know about, and they cover jobs that used to require a designer.

Blending three photos

You can upload up to three photos in one request, and Kandinsky combines them into one image, taking what it needs from each. For example: your photo, a photo of your book, and a photo of an interior. Ask: "Combine these photos into one image: I'm sitting in this interior holding this book." People use this for announcement covers, virtual outfit try-ons, interior visualizations, and product cards in different settings.

Style transfer

Upload a sample image and ask for a new one in the same style. This fixes the eternal mismatch problem: every cover in your channel looks different, and the feed looks like a patchwork. Make one cover you like, then ask: "Draw a cover about… in the same style as this image." Your feed starts to look like a brand right away.

Restoring old photos

Kandinsky 6.0 restores old pictures: it removes scratches and creases, brings back sharpness, and colorizes black-and-white photos. For a personal blog, that's a ready-made warm post: your parents when they were young, your first day of school, your first office. For a business, it's photo restoration as a service, or as a lead magnet if you're a photographer or work with family history.

The request is simple: "Restore this photo: remove the scratches and spots, make it sharper, colorize it in natural colors. Don't change the faces." If the colors come out too bright, add "muted colors, like old color film."

Reference search

If Kandinsky doesn't know what an object looks like, it finds reference images on its own and leans on them. Sber calls this Image RAG. For you, it means requests about specific places, objects, and things the model hasn't seen before now come out noticeably more accurate.

Section 10Video with sound in Kandinsky 6.0 Video

In October 2026, Sber added Kandinsky 6.0 Video to GigaChat, a model that makes video with sound built in. According to Sber, it's the first Russian AI model that can do this. A clip runs up to 5 seconds, and you pick the quality: SD (faster), HD, or Full HD. The sound covers speech and lines from characters, background music, ambient noise like footsteps and wind, and sound effects.

Five seconds sounds short, but for social media it's exactly what you need for a reel intro, a moving cover, or a transition between pieces of a video. A long video is assembled from short shots in the edit anyway.

Two ways to make a video:

• From text. Describe the scene, the motion, and the sound. "Make a video: a cup of coffee on a windowsill, steam rising, snow falling outside, the camera slowly moves closer. Sound: quiet jazz and street noise."
• From a starting frame. Upload an image (your own or one made in Kandinsky) and describe what in it should start moving. This gives you more control: you see in advance what the frame will look like.

A convenient order for video: first get the picture right in Kandinsky 6.0 Image, then bring it to life as a video. Images are free and fast, so that's where you try out versions. Video has limits, so it gets the finished frame.

What not to expect. Complex human actions (dancing, sports, hand gestures) come out worse in five seconds than the movement of nature, light, and camera. If a character needs to speak, keep the line short, 5-8 words, or the speech won't fit.

If you need longer clips with complex cuts, see my breakdown of Kling, and for a general overview of video services, read AI video from text. To animate a photo you already have, there's a separate guide on how to animate a photo with AI.

Section 11Text on images without the pain

Kandinsky 6.0 writes text on images better than earlier versions. A short word or two in large letters often comes out clean. But a six- or seven-word post headline is a lottery: one letter disappears, another doubles, and the image is no longer publishable.

The approach that works:

• In the prompt, ask for empty space for text: "empty space on the right for a headline," "clear sky on top with no objects."
• Add "no text or lettering" so Kandinsky doesn't put random letters on a mug or a sign.
• Add the headline afterward in any editor: Canva, your phone's Photos app, Figma, even the built-in story editor in Telegram.
• If you need one short word in the frame (SALE, LIVE, NEW), try asking for it directly: "a sign with the big word LIVE." Write the word in capitals and in quotes. If it doesn't work in two tries, add it by hand.

If you make lots of covers every week, there's a faster way than adding text by hand. Make the picture in an image generator, and have Claude assemble the banner from a template with your fonts and colors. I explain how that works in my guide to the AI banner maker.

Section 12Let Claude write your Kandinsky prompts

The hardest part of working with any image AI is figuring out what exactly to draw. You stare at an empty field, type "cover about burnout," and your imagination runs out. That part is easy to hand off to a text AI. I use Claude: it understands what a post is about and suggests visual metaphors you won't come up with for the tenth time in a week.

The setup is simple. You give Claude your post text or topic, it suggests three to five image ideas and writes ready prompts for Kandinsky. You copy the best one into GigaChat or whatever generator you use.

In five minutes you have a series of covers tied to the meaning of your post. Even better: ask Claude once to remember your cover style (colors, mood, what should never appear) in a project, and from then on just send topics. I explain how to set up such a project in my guide to Claude prompts. If you're outside the US and want to know how to subscribe, see how to pay for Claude.

No Claude handy? GigaChat can come up with prompts for Kandinsky itself. Type: "Suggest 4 image ideas for a post about… and draw the first one right away." Its ideas are simpler and more literal, but good enough to start.

Section 1310 ways to use it for sales and content

A picture for its own sake does nothing for you. Here's where Kandinsky, or any good image generator, actually saves time and money for everyone who sells services, courses, and consulting.

1. Telegram post covers. One cover in a consistent style for every post. Your feed starts to look like a magazine, and posts get forwarded more often.
2. Live session and webinar announcements. An atmospheric picture of a stage or a desk, with the headline and date on top. Five minutes instead of an order to a designer.
3. Lead magnet and guide covers. "An open book with paths growing out of it" sells a guide signup better than a screenshot of a document.
4. Ad creatives for Telegram and Meta. Quickly make five different versions and test which one brings leads cheaper. Free generation is the big plus here: trying options costs nothing.
5. Photos for your site and price list. An office interior, a practitioner's workspace, the feel of a studio, when you don't have your own photos yet or they don't fit the style.
6. Cleaning up your own photos. Remove clutter from the background of your avatar, channel cover, or About page photo.
7. Course lesson illustrations. Metaphors for hard topics: "a wall where one row of bricks sticks out," "scales tipping to one side." Lesson slides come alive.
8. Product cards. Your product in different settings via the three-photo blend. More on product cards in my guide to AI for marketplaces.
9. Moving reel covers. An image brought to life in Kandinsky 6.0 Video as the first five seconds of a reel.
10. Warm personal posts. A restored old photo and the story behind it build reach and trust better than any ad.

The tip that changes the look of your feed the most: create one "style tail" for your prompts and add it to the end of every request. For example: "soft daylight, beige and olive tones, minimalism, realistic photo, 16:9 format, empty space on the right." All your covers instantly become one series, and people start recognizing you in the feed by the picture before they even read the channel name.

Section 14When something goes wrong

If the result is completely off and edits don't help, open a new chat. In a long conversation GigaChat sometimes clings to early details and drags them into new images. A clean chat is a clean slate.

Section 15Kandinsky vs Nano Banana, Midjourney, and ChatGPT

Kandinsky isn't the only option, and for some tasks another tool is the better pick. For a US or European reader the real question is simple: is Kandinsky worth the extra effort next to tools you can already use? Here's a short comparison of the popular options in 2026. I'm leaving out exact prices since they change often; check the current plan pages before you subscribe.

How to choose without overthinking it. You're in the US or Europe and need work images today: start with Nano Banana or ChatGPT. You need an artistic look for a course or book cover: Midjourney. You do a lot of editing on photos with your face: Nano Banana. You want to generate in bulk on your own hardware, or build images into your own product: the open Kandinsky models are a serious candidate. You work with a Russian-speaking audience, or GigaChat lets you in: Kandinsky is free and fast. A detailed comparison of all the options is in AI image generators in 2026, and a general toolkit for different tasks is in the best AI tools of 2026.

The working pair for most tasks: one image generator for quick covers and photo edits, and Claude for all the text around them, prompts included. That pair covers about 90% of image work for a blog and ads.

Section 16Pre-publish checklist

Before you put an image in a post or an ad, run through this list. Thirty seconds that save you from awkward comments.

• Hands and fingers are fine: five each, no extra joints.
• No random letters on mugs, signs, or books.
• Faces don't look plastic: if the skin is too smooth, ask for "natural skin texture."
• The format fits the platform: 16:9 for a post cover, 9:16 for stories and reels, 1:1 or 4:5 for the feed.
• There's room for the headline, and it reads on a phone.
• The image connects to the meaning of the post; pretty alone isn't enough.
• The style matches your channel's previous covers.
• No other brands' logos, recognizable characters, or people's faces without consent.
• If it's an ad, you've checked the current terms of use.
• The image is downloaded at full size, not screenshotted from the chat.

Section 17The whole path in a minute

Start today with one action: take the topic of your next post, paste it into the "Therapist's post cover" prompt from Section 07, swap the objects for your niche, and send it to your image generator. In a minute you'll have your first cover. After a week of covers like that, you'll have a recognizable style in the feed.