The short answer
Yes, AI models draw from a description in Russian and other languages. You write in plain language what you want to see, for example «a ginger cat sleeping on a windowsill, snow falling outside», and the model creates the image.
But for the result to match what is in your head, it is worth knowing two features: how models understand non-English text and what happens to Russian text inside the picture.
Feature 1. Models understand English more precisely
Generation models were trained mostly on English-language image descriptions. So they understand English more precisely, while another language is effectively «translated» internally, and details sometimes get lost in that translation.
In practice:
- Simple descriptions («a dog on the beach at sunset») work perfectly in Russian.
- Complex descriptions with many objects, their arrangement and details miss more often in Russian: a forgotten object, mixed-up colours, the wrong arrangement.
A working compromise: write the description in your own language and add technical terms in English. Models read them more reliably:
- shot type: close-up, medium shot, wide shot;
- light: backlight, golden hour, soft light;
- style: film grain, 35mm, studio lighting.
Example: «a ginger cat sleeping on a windowsill, snow outside, close-up, soft light, film grain». This is clearer to the model than a description entirely in another language, and it does not require translating everything.
Feature 2. Russian text in the picture
A separate and very common pain. You ask for a picture with the caption «20% off» in Russian or a cafe sign, and you get letters that look Russian but have extra elements, flipped characters or are simply a mess.
The reason is that to the model letters are a pattern, not meaning, and there was several times less Cyrillic than Latin script in the training data. The mechanism is covered in detail in the article on why AI draws bad hands: text and fingers share the same nature of the problem.
The solution: generate the image without text and with empty space for the caption, and add the text itself in any graphics editor. It is faster, cheaper and guaranteed readable. Write exactly that in the description: «no captions», «free space at the top for a headline». For short captions there is an exception: Nano Banana Pro writes Cyrillic cleanly, as covered in the article on marketplace product images.
How to write the description
A good description is built from parts:
- Who or what is in frame. Specifically: not «a woman» but «an elderly woman in a knitted sweater».
- What is happening. «Reading a book».
- Where. «In an armchair by the fireplace, a cosy living room».
- Light. «Warm light from the fireplace, soft shadows».
- Style. «Photograph, film grain, muted colours».
- Format. Set the aspect ratio before generation: 1:1 for a post, 9:16 for stories, 16:9 for a banner.
What to avoid: the words «beautiful», «masterpiece», «8K», «ultra-realistic». They give almost nothing and make the picture averaged and glossy. More on this in the guide to prompts: it is about video, but the principles for images are the same.
Which model to choose
| Task | Model | Price on KADR |
|---|---|---|
| A realistic photo from a simple description | Flux.2 Pro | ~6 rubles |
| A complex description with many details | GPT-5 Image | ~46 rubles |
| The same, cheaper, with a choice of quality | GPT-5.4 Image 2 | from ~3 rubles |
| Change your own existing photo by description | Nano Banana Pro | ~25 rubles |
| A detailed picture in 4K | Seedream 4.5 | ~8 rubles |
| Graphics, icons, illustrations in one style | Recraft v4.1 Pro | ~38 rubles |
This is not a public offer. KADR prices, October 2026.
The roles in brief:
- Flux.2 is photorealism from a description, cheap.
- GPT-5 Image understands complex descriptions best, which is exactly where a non-English prompt more often lets other models down.
- Nano Banana Pro edits an existing photo with words: «change the background», «add snow outside the window».
- Seedream and Recraft are for detailed pictures and design.
If the result is not right
Common causes and what to do:
- The model ignored a detail. Simplify the description or move the important part to the beginning. Add technical terms in English. For complex scenes try GPT-5 Image.
- The picture is too «glossy» and faceless. Remove empty words, add specific light, lens and small imperfections.
- Bad hands or extra fingers. Hide the hands from focus or regenerate only that area.
- Garbled Russian text. Generate without text and add the caption in an editor.
In short
AI models create images from a description in Russian very well, especially simple scenes. For complex descriptions add technical terms in English or use models that understand long prompts better, such as GPT-5 Image. Russian text inside the picture comes out badly on most models, so add captions separately. A picture costs from 3 rubles, so finding a good description is quick and cheap.
