The main mistake
People write to a video model «make an ad for my cream» and get 6 seconds of beautiful but meaningless motion. The model does not know what a good ad is. It does what is described in the prompt, one scene at a time.
So an ad video is made the way it is in ordinary production: script first, then shooting scene by scene, then editing. The only difference is that generation replaces the «shoot».
Video structure
15 seconds (Reels, stories, video cover)
| Time | Scene | Task |
|---|---|---|
| 0 to 3 s | Hook | Stop the scroll: an unusual shot, a problem, motion |
| 3 to 10 s | Product in action | Show the product and how it solves the problem |
| 10 to 15 s | Benefit and call to action | The main idea and what to do: «Find it on Wildberries» |
30 seconds (feed ads, presentation)
More scenes are added: the problem before the product, 2 to 3 benefits as separate shots, social proof (a review, the number of buyers).
Step 1. The script
For each scene you need to describe:
- What is in frame: the product, a person, the surroundings.
- One action: «a hand puts a mug on the table».
- One camera move: «slow push-in».
- Light and mood: «morning light from a window, warm tones».
- On-screen text: you add it in the edit.
It is convenient to ask a text AI for the script. A request template for ChatGPT, Claude or DeepSeek:
You are an ad scriptwriter. Product: [description]. Audience: [who buys it]. Main benefit: [one phrase]. Write a script for a 15-second video of 3 scenes. For each scene: timing, what is in frame, one action, one camera move, light, on-screen text. Then write a prompt in English for a video AI for each scene.
On KADR this step is automated: the «Product ad» mode comes up with a presentation script from the product photo and description and gives you ready prompts with settings.
Step 2. Generating the scenes
The product always comes from a real photo. Use image-to-video mode: the product photo becomes the first frame of the scene. If you generate the product from scratch, the model will draw a similar but different one.
Drafts first. Run all the scenes on a cheap model at 480p to check whether the script works. Then the final on a good model.
Turn sound off if music and voice-over will go on top. On most models it is half the price. Model sound is only needed if a person speaks in the frame.
Example prompts for scenes:
- Hook: «Extreme close-up, coffee pours into the thermo mug in slow motion, steam rises, dark background, dramatic side light».
- Product in action: «Medium shot, woman puts the thermo mug into her backpack and walks out of the apartment, morning light, handheld camera».
- Finale: «Slow push-in on the thermo mug on a wooden table, soft warm light, clean space on the left for text».
More on camera language and prompt structure: how to write prompts for video.
Step 3. Which model to choose
| Video task | Model |
|---|---|
| Product close-up, light, texture | Veo 3.1 or Veo 3.1 Fast |
| People using the product | Kling v3 |
| A story in one piece of up to 30 seconds | Seedance 2.5 |
| A series of videos for the whole range | Veo 3.1 Lite, Seedance 1.5 Pro |
| A talking hero tells about the product | Veo 3.1, Kling v3 with sound |
More on each: which AI to choose for video, Seedance 2.5 review.
Step 4. Editing
- Music sets the rhythm. Pick the track first, then cut the scenes to its beats.
- On-screen text is large and short. Most people watch without sound.
- Logo and call to action at the end, not at the start: the first seconds are for the hook.
- Cut what is unnecessary. A beautiful scene that does not sell leaves the video.
What it costs
KADR prices, October 2026, without regenerations:
| Option | Contents | Cost |
|---|---|---|
| Budget | 3 scenes × 6 s, Veo 3.1 Lite, 1080p, no sound | ~152 rubles |
| Mid-range | 3 scenes × 6 s, Veo 3.1 Fast, 1080p, no sound | ~304 rubles |
| People in frame | 3 scenes × 6 s, Kling v3, 1080p, no sound | ~337 rubles |
| One 30-second piece | Seedance 2.5, 1080p, sound included in the price | ~4242 rubles |
This is not a public offer.
Multiply the real budget by 2 to 3: a scene with a person rarely works on the first attempt. Even with this margin it is cheaper than one day of shooting.
Moderation and refusals
Models may refuse if there is a real person, a well-known brand or someone else’s character in the frame. Seedance is especially strict about real faces. Solutions: switch models or use a generated person instead of a real one. And one rule: do not use real people’s faces without their consent.
The «as if from a blogger» format: UGC video with AI. Holiday ads: a New Year video for business.
In short
An AI ad video is a script of 3 to 5 scenes, generating each scene from a product photo, and editing with music and text. Do drafts cheaply at 480p, the final on a model suited to the task, and turn sound off if you add music. Always take the product from a real photo.
