The first time I tried this, I made the classic mistake. I wrote a dramatic prompt, got a pretty clip, and only then noticed that the product label had softened into something generic. That clip could not sell anything because it was no longer the product. So I rebuilt the workflow around one rule. The product photo is the source of truth, and the model is only allowed to change context and motion.

Now I can turn a single clean product shot into testable ads for TikTok, Instagram, and marketplace listings. This is the process I follow, including the checks that keep the product recognizable and the credits under control.

Prepare the product before prompting

I choose the sharpest image available. The product fills enough of the frame to be readable, the label is straight, and there are no distracting props. If the source photo has a watermark, price sticker, or foreign logo, I clean it first. AI will happily preserve a mistake.

Then I write down three facts the model must not change. Shape. Color. Label position. If the bottle has a matte finish, I write matte. If the handle is on the left, I write left side. This feels obvious, but it is the difference between a product video and a fictional product video.

Write the ad as a one-second decision

Before prompting, I choose the hook. There are three versions I use most. A problem hook starts with the pain and introduces the product. A use hook starts with hands or context. A result hook starts with the finished outcome and explains backward.

Then I write a prompt with one action only.

Use the product from the reference image. Keep label, color, and shape unchanged. Show a hand placing the product on a wooden bathroom shelf, soft daylight, slow push-in, commercial product video

If I need a before-and-after, I make two clips instead of asking one prompt to perform a scene change. Separate clips are easier to edit and much safer for product consistency.

Pick the right model for the job

I use free Agnes Video V2 for a rough composition test. If the framing works, I move to a production model. Seedance 2.5 is strong for expressive motion. Kling v3 Omni is my default when I need a controlled commercial shot. It costs 39 credits per second without audio and 49 credits per second with audio. Veo 3.1 gives the most cinematic feel, but it is expensive enough that I save it for an approved concept.

The model choice is not about finding one winner. It is about matching the risk. If the product must stay perfect, I choose stability over spectacle.

Aspect ratios for each destination

For TikTok, Reels, and Shorts, I generate 9:16. For Instagram feed and marketplace slots, 1:1 is safer. For YouTube or a website hero, 16:9 gives room for context. I do not rely on blind cropping. A centered product in 16:9 can get cut awkwardly in 9:16, so I adjust the framing for each version.

For 9:16, I move the product closer to the middle and keep the action visible in the lower two-thirds. For 1:1, I leave breathing room around the label. For 16:9, I allow a wider background without letting props compete with the product.

Check the output like a buyer

I watch the clip three times. First, I ask whether I would know this is the same product. Second, I ask whether the action is clear without sound. Third, I ask whether the first second tells me why to keep watching. If any answer is no, I fix that specific problem instead of regenerating blindly.

Common failures and fixes are simple. If the label warps, reduce motion and switch to image-to-video. If the background overwhelms the product, write static background, subject only moves. If the output feels slow, cut the first half second in editing. If the product appears too late, regenerate with product visible in the first frame.

Add claims and CTAs in the editor

I do not ask the model to generate offer text or compliance claims. AI typography can shift, and legal wording needs to stay exact. I overlay Free shipping, Dermatologist tested, or whatever the campaign actually promises, then keep the text within safe margins.

This also makes localization easier. The same video can serve different markets with different overlays. The visual remains stable while the message changes.

Build a reusable template

Once one product works, I save four things. The source photo requirements. The hook type. The model and settings. The final prompt. For the next SKU, I replace the product-specific words and keep the rest. This turns AI generation from a creative lottery into a production system.

A practical starting budget is one free draft, one approved still, and one short production clip. If that clip proves the concept, I make 9:16 and 1:1 versions. If it does not, I adjust the prompt instead of scaling a bad idea.

How to Create AI Product Videos for E-Commerce Ads