Instagram and TikTok feeds are packed with product videos that nobody ever filmed: they were made by AI. A video like this comes together in one evening from supplier photos - no studio, no sample, no actors. Below is the workflow people use right now: which tools, in what order, what to write in your prompts and where most people slip up.

Step 01Understand what goes into a video like this

Nearly every viral AI product video follows the same four-part formula:

  1. A scene image. The product isn't on a white background but in real life: in a kitchen, in someone's hand, on a table by the window. An AI image generator creates it from the supplier's photo.
  2. Animation. A video model turns that image into a 5-8 second clip: the camera glides, a hand picks up the product, steam rises from a mug.
  3. Voice or text. An AI voiceover, or simply large on-screen text over a trending sound.
  4. Editing. 3-5 of these clips are stitched together into a 15-25 second video in CapCut.

The golden rule: don't animate a flat white-background photo directly. Video models struggle with product listing images: the first frame comes out frozen and the motion looks unnatural. Build the scene first, then bring it to life.

Step 02Create a product scene from the supplier's photos

Grab the 2-3 sharpest product photos from your supplier (AliExpress, CJ, 1688): front, side and in hand. The more accurate the shape and color in the source image, the less the AI will make up.

Scenes are made in image models that can keep the product true to a reference. The most popular one right now is Nano Banana (Google Gemini): upload the product photo and describe where it should be. Others work too: GPT Image in ChatGPT, Flux Kontext, Seedream.

Sample prompt (write in English - it's more accurate):

Place this exact product on a wooden kitchen counter, morning sunlight from the window, a woman's hand reaching for it, shot on iPhone, realistic, vertical 9:16. Keep the product shape, color and logo exactly the same.
  • That last sentence is a must: without it, the AI "improves" the product, and it no longer matches what the customer will actually receive.
  • Make 4-6 different scenes: a close-up, the product in action, a "before and after" result, the product in someone's hand.
  • Go vertical 9:16 from the start - for Reels, TikTok and Stories.

Step 03Animate the scene with an AI video model

Upload the finished scene as the first frame (image-to-video) and describe the motion. The most commonly used models:

  • Veo 3.1 (Google, in Gemini and Flow) - the most realistic product shots, and it can add sound right away. Great for "premium" close-ups.
  • Kling - cheaper, handles people and hands really well, and lets you set a first and last frame so the model builds the transition between them. Handy for "before and after".
  • Sora 2 (OpenAI) - better for lifestyle scenes and mood than for precise product detail.

If you don't want to sign up for every service, there are platforms that bring all these models together in one place: Higgsfield, Freepik, Picsart Flow. You pay per generation and pick the model for each shot.

Sample motion prompt:

Slow camera push-in, the hand picks up the product and turns it to show the side, soft natural light, realistic handheld phone footage, 6 seconds.
  • One action per clip. "The hand picks it up, opens it, pours and smiles" in a single clip almost always falls apart.
  • Generate 3-4 versions of each clip and keep the best one: duds are normal with AI.
  • Check the product in every frame: extra buttons, a "melting" logo, six fingers - that clip goes in the trash.

Step 04Add a voiceover or on-screen text

There are two formats that work:

  1. No voice. A trending sound plus large text: the problem in the first frame ("Tired of scrubbing your stovetop?"), the solution, the result. The simplest and fastest option.
  2. Voiceover. Voice a 40-60 word script in ElevenLabs or right inside CapCut. Pick a natural, conversational voice, not a "radio announcer" one.

The "real person talking about the product" format (UGC) is made with AI avatar tools: Arcads, HeyGen, Creatify. You write the script, and the avatar says it to the camera while holding the product. These videos perform well in ads but cost more than regular generations.

How to write a script people watch to the end:

  • The first 2 seconds should be the weirdest, most useful or most unexpected thing: a result, a problem, a number.
  • One benefit per video. Nobody will remember five benefits in 20 seconds.
  • Short sentences, 7 words max.

Step 05Put the video together in CapCut

  1. Clip order: hook (problem or result) - product in action - close-up - result - call to action.
  2. Length: 15-25 seconds. Each clip 2-4 seconds, no longer: fast cuts hold attention.
  3. Subtitles - CapCut's auto captions, large, in the center of the screen: most people watch with the sound off.
  4. End with an action: "link in bio", "comment a keyword". If DM replies are automated with a bot, people get the product link instantly.

Make 3-5 different videos per product right away, each with a different hook and first frame. You can't predict which one will work - that's normal, everyone tests this way.

Step 06Don't break the rules, or you'll lose money

  • The video must match the real product. If the AI "added" a feature that doesn't exist, or made the product look bigger and pricier, customers will demand refunds and banks will open disputes. For Stripe and PayPal, lots of disputes means frozen funds and a closed account.
  • The "Made with AI" label. Meta automatically adds an "AI info" label to realistic AI content and asks you to disclose it yourself. It doesn't hurt reach, but passing off an AI person as a real customer giving a review is a direct violation of ad policies.
  • No other brands or faces. Don't use well-known brand logos or celebrity faces in your prompts - the video will be taken down, and your ad account may be restricted.
  • Claims. No "cures", "lose weight in a week" or "guaranteed results" - videos like that won't pass ad review.

SummaryA quick one-evening workflow

  1. Pick a product and download 2-3 sharp photos from the supplier.
  2. Create 4-6 scenes in Nano Banana with the phrase "keep the product exactly the same".
  3. Animate each one in Veo 3.1 or Kling, 3-4 versions each, and keep the best.
  4. Add on-screen text or an ElevenLabs voiceover.
  5. Assemble 3-5 videos of 15-25 seconds in CapCut, each with a different hook.
  6. Check that the product in the video matches the real one.
  7. Post them and watch which video gets the most completed views - that's the one you scale.

AI prices and model versions change every couple of months, but the workflow itself - scene, animation, voice, editing - stays the same. Learn it once, and you'll be able to make a video for any product in a single evening.

Message us on TelegramWe'll help you launch your store
Enjoyed the article? Share it