What an AI product film actually costs (we made one without a studio, a crew, or the product)

We built a product film for a camera nobody has ever manufactured. It comes apart as you scroll: the lens leaves first, then the sensor, the boards, every last screw. You can scroll it apart yourself on our site.

This was a test, not a client job. We wanted to know what one of these actually costs to make now, so we built one properly and measured everything about it.

There was no studio booking. No crew. No lighting rig. No product sample couriered to a photographer, because there is no product. We are publishing what it took, in the same spirit as our bill for the scroll film on thesocialtarget.com/world: a receipt, not a case study.

Three questions sit underneath most of the search traffic on this subject, so we are answering them before anything else: how much does a product video cost, can AI make a product video, and how do you make an exploded-view animation.

How much does a product video cost?

There is no honest single number, and we are not going to invent one. What we can price is our own: a complete exploded-view product film delivered as 149 rendered frames weighing 9.3 MB, with no studio, no crew, and no physical product involved at any stage.

The reason most published answers to this question are useless is that they price a production line rather than an asset. A day rate, a crew size, a kit list, a grade, an edit. Those are the costs of a particular METHOD of getting a film, and they are the costs that have just stopped being compulsory.

We are not quoting you a market average, because we could not verify one we would be willing to defend in print. Nearly everyone publishing a number for this sells the service. We do too, which is exactly why we would rather show you our own bill than someone else's estimate.

So here is the useful version of the question. Not "what does a product video cost", but "which parts of the old bill do I still have to pay?" On this job, the answer was: the thinking, and nothing else. Every line item that used to dominate the quote was a line item for capturing footage of a real object, and there was no real object.

Can AI make a product video?

Yes, for a specific and narrow job: a controlled, mechanical, single-subject sequence like a product coming apart or going together. It is far less reliable for anything involving people, hands, brand-accurate packaging, or a product that has to match a real one your customer already owns.

The method that works is not "describe a commercial and receive a commercial." That is still the version that produces something almost right and completely unusable.

What works is locking the endpoints. You render the start and the end as still images first, because a still is the thing you can control, judge, and re-do cheaply. Then you ask a model to travel between two frames you have already approved. The motion is the only part being generated, and it is bounded on both sides by pictures you already signed off.

Our own page describes it in three steps, and they are worth reading as a method rather than a boast. Step one: the assembled hero and the fully exploded view are generated as stills first, same camera, same light. Step two: a video model travels between the two locked frames, with no animation software, no 3D scene, and no camera crew. Step three: the result is split into frames and painted to a canvas by scroll position.

That is the whole trick. Two pictures and a guess between them, where the guess is fenced in on both sides.

How do you make an exploded-view animation?

Traditionally, from a 3D model: someone builds or receives CAD geometry, separates the parts along their axes, animates the separation, lights it, and renders it. The route we took skips the geometry entirely: render two stills, interpolate between them, then drive the playback with the visitor's scroll instead of a play button.

The conventional route is genuinely good and it is genuinely expensive, and the expense is mostly upstream of the animation. You need the CAD files, or someone to model the product from scratch, and modelling a product accurately is a specialist job priced like one.

The route we took has a different shape. There is no geometry at any point, which is precisely why the subject is a camera that does not exist: nothing had to match a real object, so nothing could be wrong. That is a real constraint, not a clever framing, and we come back to it in the limits below.

The last step is the one people underestimate. Splitting the film into individual frames and painting them onto a canvas driven by scroll position does two things at once. It puts the viewer in control of the timeline, which is why a scroll-driven assembly reads as mechanical rather than as an advert. And it removes the video file entirely, which turns out to matter more than it sounds.

What our own film actually cost

Every number here was measured on disk, and you can check most of them yourself by loading the page.

Worth setting beside our other showcase page for contrast. It is 64 MB and runs on about 1,000 lines of JavaScript, as we published in our article on when website animation earns its keep. This one is 9.3 MB on a desktop and 1.6 MB on a phone. The plainer method turned out to be cheaper AND lighter, which is not the trade people expect.

Why it is 149 pictures and not a video

The page contains no video file at all. It is a sequence of JPEGs painted onto a canvas element, one per scroll position. That sounds like the cheap version and it is actually the reliable one, because scrubbing a video backwards and forwards is a decode-heavy operation that phones ration.

We learned this the expensive way on the /world page, and wrote it up at the time: iOS limits how much video a page may decode at once, so an effect built on a desktop can stutter badly exactly where most of your visitors are standing.

Frames sidestep the problem. Painting the next picture onto a canvas is something every device can do at any speed, in either direction, with no decoder involved. Scroll down and it comes apart, scroll up and it reassembles, and it does that identically on a five-year old phone.

The mobile version is not the desktop version shrunk. It is a genuinely separate, shorter frame set, chosen once when the page loads, so a phone never downloads the 9.3 MB desktop sequence at all. A phone gets 1.6 MB. Until the choice is made, the visitor sees a poster image that is byte-identical to frame one, so the swap is invisible.

None of that is exotic engineering. It is a sequence of pictures and a scroll listener, and the visitor cannot tell the difference from something far more expensive.

What this method cannot do

This is the section most articles on this subject leave out, so it is the longest one here.

It cannot match a product that actually exists. Our subject is a camera nobody ever built, and that is not a cute detail, it is the load-bearing assumption. The moment the film has to depict YOUR product, with your proportions, your materials, your logo in the right place and your actual screw count, you are asking a generative model to be accurate rather than plausible. It is good at plausible. Accuracy is where this gets expensive again, and where CAD-based rendering earns its money back.

It cannot be trusted with text or branding. Anything with a legible logo, a label, a regulatory mark, or packaging copy is a risk on every frame. 149 frames is 149 chances to be subtly wrong.

It cannot do people or hands well enough to ship. A product film with a human using the product is a different job with a different failure rate. Everything in this article applies to an object alone in a void.

It is not one prompt and done. The controlled part of this method is the two stills, and those are iterated like any other piece of art direction. What collapsed in price was the production line, not the number of judgement calls.

It cannot be indexed as content. A canvas full of painted pictures carries no text for a search engine or an answer engine to read. Our page is deliberately allowed into search while the rest of our lab pages are not, but what earns it a place in search results is the words around it, not the film. Any product film, made any way, is an asset on a page and never a substitute for one.

And it does not tell you what to show. This is the real limit and it does not have a technical fix.

So what should you budget for now?

The old constraint was the production line. If a product film cost four figures and you did not have four figures, you did not have a product film, and that was the end of the conversation. That wall has fallen over.

What is left standing is the part nobody could ever buy: knowing what is worth showing. Our page ends on the line the whole exercise proved, and we will let it stand as the conclusion here too. The expensive part was never the pixels. It is knowing which sixty seconds of engineering to show.

So move the budget. Stop pricing the shoot and start pricing the decision: which single mechanism inside your product would change a customer's mind if they could see it, and what does it have to look like to be believed. That is the brief now, and it is a strategy question wearing a production question's clothes.

If you want a straight answer on whether your product suits this treatment or still needs a real camera, tell us about the product through our intake form and we will tell you plainly which of the two it is, and why.