AI product video: what it does well and what still needs a camera

The part people talk about less is what happens when twelve of those clips sit next to each other on a category page, and the red of your bag is a shade deeper in four of them. Nothing in that set is bad. The set is wrong.

This article is about where generated video is genuinely good enough today, where it is not, and how to decide that per product instead of per opinion. You will not find a tool comparison here, for the same reason we gave in why we do not publish a tool list: any list we wrote down today would be wrong within six months, and which tools we use is a choice we make per assignment.

What generated video does well

Four things, and they are not small.

Motion from a still. A product that turns, a slow push towards a detail, a background that drifts gently while the product stays put. This is the most reliable category by a distance, because the product itself is already fixed by your own photograph.

Context, swapped. The same product in a summer setting and a winter one, in a Scandinavian room and a German one, on a terrace and in a shop. For a brand that sells into several countries this is the part that changes the arithmetic, because the alternative is travelling.

Volume. Fifteen variants of the same fifteen seconds, for a campaign you want to test before you commit to it. No crew has ever been able to do that economically.

Formats. Vertical, square and wide from one source, without a second edit.

Where it still falls over

Length. Past roughly ten to fifteen seconds, a generated clip starts to wander away from the thing it was given. Short is not a stylistic preference here, it is where the technology is reliable.

Small type and marks. Ingredient panels, model numbers, a logo on a curved surface. These degrade in ways that are easy to miss on a phone and obvious on a desktop.

Colour, held across clips. This is the one that catches people out, and it is worse in video than in stills. A still that comes out half a shade off can be corrected once and it is done. A clip is a few hundred frames, and the drift lives inside the motion, so correcting it is not a colour pass, it is a regeneration. If your brand leans on an exact colour, assume every clip needs checking against the real thing rather than against the previous clip. The same compounding problem shows up in stills once a range grows, which we wrote about in what breaks first when a range grows.

Materials with behaviour. Fabric falling, liquid pouring, transparent glass, high-shine metal. The model does not know what your material weighs.

Anything that demonstrates a mechanism. If the point of the clip is that the lid closes a particular way, or that the drawer runs soft, that has to come from the real product. A generated version will produce something plausible, and plausible is exactly the problem.

The question that settles it: is the clip load-bearing?

Here is the test we use internally, and it is more useful than any category list.

A clip is load-bearing when the buyer uses it to decide something factual. How big is it really. How does it open. How does the fabric fall. What is the finish like up close. If the answer in the clip is wrong, the buyer finds out when the box arrives, and you pay for it in a return and a review.

A clip is not load-bearing when it sets a mood, establishes context or fills a feed. It still has to be on brand and it still has to be honest, but nothing in it is a claim about the product.

Load-bearing clips come from the real product. Everything else is open to generation. That single line removes most of the argument from the discussion, and it survives the next model release, which a category list does not.

The part that matters if you sell through other people

Most of what is written about AI video assumes one brand with one webshop. That is not our reader. If you sell through dealers, retailers and marketplaces, a clip does not live on your product page. It gets downloaded, re-uploaded, trimmed, and put on sixty sites you do not control.

Two consequences, and neither is obvious until it happens.

First, a mistake in a clip is much harder to take back than a mistake in an image. An image you can replace in your library and in a dealer feed. A clip has usually been re-uploaded to a platform by the dealer, where it now has its own identity, and your replacement does not reach it.

Second, the clip that travels furthest is the one with the least in it. No seasonal set dressing, no on-screen price, no local claim. Make the neutral version the one you distribute, and keep the dressed-up variants for channels you run yourself. The same logic applies to stills, and it is one of the reasons one library everyone works from beats sending files around.

What we check before a clip leaves our hands

We produce video the same way we produce stills: generated, with a creator from our own team in the loop. Four checks happen every time, and they exist because each one caught something real.

1. Three frames against the packshot. First frame, middle, last. Not against each other, against the actual product photograph. Drift is invisible frame to frame and obvious over fifteen seconds.

2. Marks at full size. Logo, model name, any printed panel, viewed at 100 percent and not on a laptop preview.

3. The physics pass. Does anything move in a way the object cannot. Handles that bend, a hinge that opens past its stop, a textile that behaves like liquid.

4. The mute test. Nearly everyone will see it without sound. If the clip only makes sense with the music, it does not make sense.

That fourth one sounds trivial and it is the one that sends most clips back. The same pattern we described in what we learned making AI product photos at volume holds for video: the technology is rarely the bottleneck, the rules you put around it are.

So where does that leave you

Use generated video for context, for variants, for formats and for volume. Use the real product for anything the buyer will use to make a decision. Keep clips short, check them against the product rather than against each other, and distribute the plain version.

If you want to work out which route fits a specific product, we wrote that up separately in which production route fits, and what a clip has to do once it is on the page in what a product page video has to do.

Want to see how we produce this, start to finish? See how it works, or try LensLab free with five credits.