Format first, then generation
The most common mistake: generate a beautiful horizontal clip and then try to fit it into a story. Cropping 16:9 down to 9:16 loses more than half the frame, the hero ends up past the edge, and the camera move drifts into emptiness.
So the first step is to decide where the content goes:
| Where | Format | What goes in |
|---|---|---|
| VK and Telegram stories | 9:16 | backgrounds, announcements, short clips |
| VK clips, Shorts | 9:16 | vertical clips |
| Feed post | 1:1 or 4:5 | covers, illustrations |
| Video thumbnail | 16:9 | the clip’s cover |
| Channel or community header | wide horizontal | banner |
The format is set before generation. All major models support it.
Vertical composition is a different composition
In a vertical frame, what works is anything that stretches upwards: a person at full height, a close-up of a face, a street receding into the distance, a bottle on a table. Wide scenes work badly: panoramas, group shots, long tables.
It also matters where the platform’s interface will be. In stories its elements cover the picture at the top and bottom, so keep the main thing in the central part of the frame and leave breathing room at the edges.
Which model for which content
Covers and illustrations from scratch. Flux.2 gives a realistic picture for just 6 rubles; it is the workhorse for volume. If the composition is complex and the description has many details, take GPT-5 Image.
Processing your own photos. Changing the background, evening out the light, giving a series of real shots one style: that is Nano Banana Pro.
Graphics and design. Highlight icons, neat graphic covers, brand styling: Seedream 4.5 and Recraft are good here; more in their review.
Short clips for stories and clips. The inexpensive Veo 3.1 Lite (about 84 rubles for 6 seconds at 1080p) and Seedance 1.5 Pro (about 136 rubles) cover atmospheric and product clips. For scenes with people and expressive camera work, Kling v3 is better.
The pipeline: a month of content in an evening
If you make one picture every day, it turns into a chore. It is much more efficient to assemble a month in one go.
1. A content plan. Write down 20 to 30 post topics. You can ask a text AI to suggest topics for your niche.
2. One style for everything. Formulate the style description once (palette, light, lens, mood) and copy it word for word into every prompt. That way the feed looks coherent rather than like a collection of random pictures.
3. Debug on one picture. Tune the prompt on one example on a cheap model. Only then run the rest.
4. Batch generation. Run all the topics in one style.
5. Text separately. Add captions in an editor: text inside a generation, especially Cyrillic, often comes out badly.
6. Clips selectively. You do not need a video for every post. Two or three clips a week for clips and stories give more reach than thirty static pictures.
How not to look like everyone else
The problem with AI content on social media is not quality but sameness: gloss, golden hour, faceless beauty. A feed made of such pictures gets scrolled past without a glance.
What helps:
- A specific style instead of «beautiful». Lens type, character of the light, palette.
- Imperfection. Film grain, slight camera shake, flares. A perfect picture reads as artificial.
- A reference instead of words. If you have a picture with the right mood, feed it as a sample.
- One style for the feed. A coherent profile looks more expensive than any single beautiful picture.
More on these techniques in the article on how to make Reels with AI and in the guide to prompts for video.
What it costs
An estimate for a month for one account (KADR prices, October 2026):
| What | Quantity | Approximately |
|---|---|---|
| Post covers on Flux.2 Pro | 30 | 190 rubles |
| Processing your own photos on Nano Banana Pro | 10 | 253 rubles |
| Short clips on Veo 3.1 Lite | 8 | 670 rubles |
| Margin for regenerations | x2 |
This is not a public offer.
In total about 2200 rubles a month including failed attempts. How else to save on video (resolution, sound, choice of model) is covered in detail in the article on how much a minute of AI video costs.
Video covers with text: a YouTube thumbnail with AI. Profile portraits: an AI photoshoot from your photo.
In short
Format for the platform first, then generation: vertical is made straight in 9:16. Images from scratch on Flux or GPT-5 Image, processing your own photos on Nano Banana Pro, short clips on the inexpensive Veo Lite or Seedance. One style for the whole feed, text separately in an editor, clips selectively. A month of content is assembled in an evening and costs a couple of thousand rubles.
