How to Generate Images with Fusion Brain (Kandinsky) for Free

November 10, 2026

Fusion Brain is a free web app from Sber that runs the Kandinsky image models — currently the 3.x line — in your browser. You type a prompt, choose a style, and get images in about 20 to 60 seconds. No account, no API key, no credit card.

That combination is rare. Most text-to-image tools gate output behind a subscription or hand you watermarks. This guide to Fusion Brain covers the actual workflow and, just as importantly, where the free version stops being enough.

What can Fusion Brain generate?

Three modes: text to image, image to image (roughly a redraw), and inpainting for editing part of an existing picture. Resolution tops out at 1024×1024 on the main text-to-image mode, with portrait 768×1344 and landscape 1344×768 available.

The underlying model is Kandinsky 3.1, released in 2024, and Sber’s earlier 2.x models remain selectable in some views. Kandinsky is trained heavily on Cyrillic-language imagery, which is why prompts about Slavic signage, folk motifs, and Russian cityscapes come out unusually coherent — and why some Western pop-culture references land a beat off.

Who should use it, and who should not?

Good fit: people who need a handful of usable images a week for slides, blog headers, or mockups, and who do not want another subscription. Students and hobbyists experimenting with prompt styles. Anyone behind a corporate firewall that blocks other image services.

Poor fit: commercial design work that demands consistent characters across many images (there is no reference-image consistency feature here), photorealistic product shots, and any workflow that needs an API. For production pipelines, something with a paid API and commercial licensing, such as Ideogram or a Stable Diffusion deployment you control, makes more sense.

Also relevant: the interface and documentation mix Russian and English. Translation is usually good, occasional labels are not.

Step 1: Open it and pick a mode

The web app opens on text-to-image by default. Learn that mode first — image-to-image and inpainting each have their own quirks that will only confuse you if you start there.

No signup is required for basic generation. That also means no saved gallery across browser sessions, so download what you like immediately.

Step 2: Write a prompt that Kandinsky understands

Prompt order matters more than prompt length. Use noun phrase first, then attributes, then style:

portrait of an old lighthouse keeper, weathered face, wool cap, soft overcast light, 50mm photograph

Versus the vague version: “a nice picture of a lighthouse man.” The second produces a generic stock photo.

Three rules that hold up in testing:

  • Name the medium. “oil painting”, “watercolor sketch”, “35mm photograph”, “isometric render”. Without this the model defaults to a glossy digital look.
  • Keep it under 60 words. Longer prompts dilute; the model weights the first phrases most heavily.
  • Use negative prompts. The field exists in advanced settings. Common entries: text, watermark, extra fingers, blurry, low quality.

Step 3: Choose aspect ratio and style before generating

The negative prompt field and style presets are behind a settings toggle. Set the aspect ratio first — regenerating at a different ratio re-rolls the whole composition, so a portrait result never resizes cleanly into a banner.

Style presets (anime, oil painting, photorealism, and so on) act as prompt modifiers. Stacking a preset with an explicit “oil painting” in the prompt tends to over-bake the style.

Step 4: Iterate with seeds, not rewrites

Each generation returns a set of variations. Keep the prompt fixed and re-run a few times before you start editing words. When one result is close, note the seed if it is exposed and lock it, then change exactly one phrase.

Editing three things at once is the most common reason people conclude the tool “doesn’t work”. You cannot tell what caused the change.

What are the actual limits of the free version?

Generations per day are capped and the queue can be slow during peak hours in Russian time zones. Commercial use of generated images is granted under the model’s licence, which you should read yourself before selling anything — the terms have changed between model versions and the English version has been revised more than once.

There is an official API, but it requires a Russian phone number for registration, which rules it out for many teams outside the region.

Quick Summary

  • Fusion Brain is a free browser app running Sber’s Kandinsky 3.1; current max output is 1024×1024, with 768×1344 and 1344×768 options.
  • Three modes exist: text-to-image, image-to-image, and inpainting.
  • Put the subject first, name the medium, and keep prompts under 60 words for the best hit rate.
  • No account is needed, which also means nothing is saved — download immediately.
  • No character consistency, no easy API access outside Russia, and no commercial-license guarantees you should assume without reading the terms.