Ecommerce Product Videos: A Practical AI Production Workflow

Product video earns its place on a page for one reason: it answers something a still photograph cannot. How the hinge locks. How the serum spreads. Whether a sixteen-inch laptop actually slides into the sleeve. Everything else — the marble surface, the slow orbit, the string swell — is decoration that costs money and proves nothing.

The economics have changed, though. Generating motion from a reference image or a written brief means a two-person team can produce a dozen product clips in the time it used to take to book a studio, hire a hand model, and wait for the edit. But generation also makes it easy to ship glossy footage that is subtly false: a bottle with the wrong cap, a zipper with invented teeth, a fabric that reads as plastic. The workflow below exists to keep the speed and remove the falsehood. It treats generation as one step in a pipeline that begins with a sales question and ends with a measured test, not with a render.

Start with the sales question, not the shot list

Before opening any generator, write one sentence per product that names the objection the video has to dissolve. Not a mood board. A sentence.

• A folding chair: buyers cannot tell whether the hinge looks flimsy, so the clip must show the lock engaging under load.
• A face serum: buyers cannot tell whether it absorbs or sits greasy, so the clip must show texture spreading on skin at close range.
• A commuter backpack: buyers cannot judge whether a laptop fits, so the clip must show the machine sliding into the padded sleeve and the flap closing.

That single sentence decides camera angle, motion, duration, and sound design. Teams that skip it end up with beautiful generica — slow orbiting shots of objects on pale stone that could belong to any brand in the category and answer nothing.

A useful sorting rule: if a still photo already answers the question, do not make a video of it. Video earns its place only when motion, scale, sequence, or texture carries information a still cannot. Write the rule on the wall, because the temptation to generate everything is strong and expensive.

Decision criteria for whether a product deserves video at all:

1. Movement — does the product do something (fold, click, pour, stretch)?
2. Scale ambiguity — do buyers misjudge size from photos?
3. Comparison pressure — is the category one where shoppers read specs before adding to cart?
4. Return risk — do returns cluster around a misunderstanding a clip could prevent?

If two or more answers are yes, video pays for itself. If none are, improve the photography instead.

The four formats worth making first

Most catalogs do not need a bespoke hero film per SKU. They need a small set of repeatable formats that share assets, prompts, and style rules — which is exactly what makes them affordable.

The hero loop

Six to ten seconds, silent by default, looping cleanly, and readable as a product within the first second. It sits in the first gallery slot or at the top of the product page. Build it as one continuous motion — a slow push-in, a quarter rotate, a reveal from behind fabric — with no cuts, because cuts inside a loop read as a stutter.

This is the hardest format to make convincing with generation, because viewers scrutinize the product closely and hold it in their hands for comparison. Generate it last, once you have learned how the model handles that particular object, and keep a real photograph as the fallback if three attempts still drift.

The feature explainer

Fifteen to forty seconds with three to five beats, each tied to a specification. Use it where buyers compare before buying: electronics, tools, kitchen equipment, outdoor gear. Because it has cuts, you can interleave generated shots with real macro photography and screen recordings, hiding the weak points of generation behind genuine product detail.

The proof clip

Vertical, eight to fifteen seconds, styled like something a customer or a creator shot on a phone. Low polish is the point: natural light, a hand in frame, a slightly imperfect composition reads as believable. Keep every claim inside what the product actually does, and never synthesize a testimonial voice you do not have. If real customer footage exists, this is the format to cut it into.

The retargeting cutdown

Three to six seconds, one idea, built for muted autoplay in front of an audience that already visited. This is where the hero loop gets reused with a different first frame, a tighter crop, and an overlay that leads with an offer or a sharper benefit. Cutdowns are nearly free once the master exists, and they usually return the most per minute of work, because the audience is warm.

Match each format to the placement it will actually run in. A forty-second explainer dropped into a retargeting slot wastes spend; a three-second cutdown on a product page looks like a technical fault.

From product truth to a defensible shot list

Generation is fast enough that the bottleneck has moved upstream. Teams that ship well spend their hours collecting references and writing shot lists, not endlessly iterating prompts.

Run an asset audit first

Gather everything you own: catalog photography, 360 spins, unboxing clips, packaging shots, lifestyle images, material close-ups, color swatch files. Then list what is missing. In practice most catalogs lack three things — a mid-motion frame, a scale reference, and a texture macro.

Those gaps define what generation should produce. A model that sees your real product photograph returns far more faithful shots than one working from adjectives alone, because it can lock shape, logo placement, and material response. If you need still plates to animate later, an AI image generator is often a better first stop than going straight to motion, since you can reject a wrong silhouette in seconds instead of minutes.

Write the shot list as a table

A workable shot list has six columns: shot number, format, what it must prove, camera, duration, and asset source. The asset source column — real photo, generated plate, screen recording, existing clip — forces the honest question of whether generation is needed for that beat at all.

Two rules keep the list realistic. First, no more than one complex camera move per shot; a push-in plus a rotate plus a rack focus is where artifacts appear. Second, every generated shot must be replaceable by a real one later. If a shot cannot be reshot practically, do not put it at the center of the page.

Prompts written as specifications

A prompt for product video is not a description of a pleasant scene. It is a specification. Six blocks in a fixed order produce predictable results and, just as importantly, make failures diagnosable.

1. Subject, with the detail that matters: material, color, and the exact part
2. Camera position and lens behavior: macro, focal length equivalent, working distance
3. Motion, stated as one verb and a speed
4. Light: source, direction, hardness, and what it does to the material
5. Environment, described with restraint
6. Constraints: what must not appear, change, or deform

The sixth block is the one teams omit, and the one that saves the most renders.

Subject: matte ceramic pour-over dripper, unglazed, fine speckle, spiral ribs visible
Camera: macro 100mm equivalent, slight high angle, 20cm working distance
Motion: one slow push-in, three seconds, no rotation
Light: single softbox from camera left, hard rim from rear right to trace the ribs
Environment: pale linen surface, single soft shadow
Constraints: keep rib count and lip thickness constant, no text, no logo, no hands, no steam flicker

Compare that with "beautiful coffee product cinematic video." The second gives the model freedom, and freedom is how products grow extra parts. When a shot fails, change one block at a time — usually constraints or light — and regenerate. Changing three things at once teaches you nothing and burns the afternoon.

Teams working across many SKUs should save a base template per category and swap only subject and constraints. Starting from a ready-made structure is faster than starting from an empty field, which is why browsing video templates before writing from scratch usually shortens the first pass. Once a template survives ten SKUs, it belongs in a prompt library the whole team can draw from.

Useful constraint phrases that prevent the most common failures: "keep label text legible and unchanged," "no hands," "product remains centered in the middle third," "no new parts, no added buttons," "no morphing at the silhouette edge." Each one targets a specific failure mode you have already seen.

Consistency rules for a whole catalog

One good clip is a demo. A catalog that looks like a single brand made it is a system. The visual constants are unglamorous: the angle of the sun, the background tone, the color temperature of white, the speed of camera movement, the length of the fade.

Define those constants once, in writing. A one-page style sheet with four lines — background tone, light direction, motion speed, color grade — removes most inconsistency before it happens, because every prompt inherits it.

The levers that do the heavy lifting:

• Reference before adjectives. Feed the model a real product image. Shape fidelity comes from pixels, not from the word "sleek."
• Reuse a seed family. Keep a shot on the same seed line and vary only the block you need, so grain and lighting do not drift between adjacent clips.
• Lock crop and horizon. Fix the product's vertical position across clips. When products hop up and down the frame during scroll, the page feels unbuilt.
• One motion vocabulary. If some clips orbit and others dolly, the grid looks assembled from different companies. Choose two moves and stay with them for a season.
• Grade at the end, in a batch. Apply one look to the whole set. Per-clip grading is how catalogs end up with eleven slightly different whites.

Review as a contact sheet: all clips as thumbnails in one view. Problems invisible in a full-screen player become obvious in a grid — mismatched backgrounds, shadows pointing different directions, one product subtly larger than its neighbors.

Assembly, sound, and the cutdown matrix

Generation gives you pieces; the edit decides whether they work. Treat assembly as its own discipline with its own rules.

• Move fast in the first second. A product clip at the top of a page has under a second to signal what the object is. Open on the product, not on an abstract texture, a logo sting, or a fade from black.
• Keep cuts short. One idea every two to three seconds is plenty. Long holds invite scrutiny of generated detail.
• Design for silence. Most product-page and social plays are muted. If a clip only works with sound, it does not work.
• Caption anything with speech. Beyond comprehension, captions are a baseline accessibility expectation for time-based media. Size them for mobile, keep contrast high, and check that overlay text avoids the platform's interface zones.
• Use sound as texture. Room tone, a cloth rustle, one clean click of a hinge. Music that swells at the reveal is the fastest way to make a generated clip feel dated.
• End on the product. In frame, undistorted, correctly lit. The final frame is the one that gets screenshotted.

Assemble in short sequences and export clean masters at the highest resolution you can afford. Cutdowns are trimmed from the master, never re-generated, so a three-second retargeting clip stays frame-identical to the hero shot it came from.

Placement	Shape	Typical length	What changes from the master
Product page hero	16:9 or 1:1	6–10s	Loop point only
Gallery slot	1:1	8–15s	Square crop, product re-centered
Feed ad	4:5	6–10s	New first frame, tighter crop
Stories / short video	9:16	6–15s	Vertical reframe plus text safe zones
Retargeting	1:1 or 9:16	3–6s	One beat, offer overlay, hard end card
Email	16:9	3–5s	Silent, small file, no tiny text

Verify current pixel and duration requirements against each platform's own documentation before shipping; the matrix above is about structure, not spec sheets. The vertical reframe is where most quality is lost. Do not simply crop the center — plan a vertical-safe composition during generation, keeping the product in the middle third and dead space top and bottom for text. If the product fills a landscape frame edge to edge, no reframe will save it.

A ninety-second QA pass

Run the same checklist against every clip before it goes near a product page. Ninety seconds catches almost everything.

• Shape check. Compare against a real photo taken at the same angle. Count ribs, buttons, seams, stitching. Confirm logo placement.
• Text check. Any lettering on the product must be legible and correct. Garbled microtext is the most obvious tell.
• Motion check. Watch at half speed. Look for warping at edges, melting hardware, flickering shadows, objects drifting between frames.
• Physics check. Liquid should fall, fabric should fold, hinges should have stops, straps should carry weight.
• Color check. Compare a frame against the swatch file, not against your memory of the product.
• Claim check. Every visual claim must match the specification sheet. A clip implying waterproofing on a water-resistant item is a legal problem, not a creative one.
• Accessibility check. Captions present, contrast adequate, no flashing sequences, no text that flashes at the loop point.
• Weight check. If a product-page clip runs several megabytes, it is costing more in load time than it earns in engagement. Compress, lazy-load, poster-frame it.

The recurring mistakes are predictable. Generating without a reference image. Letting the model invent hands. Using a wide shot to hide detail problems. Showing the product from an angle a customer never sees in real life. Shipping a clip whose loop point jumps. All six are cheap to fix at the shot-list stage and expensive to fix after publishing.

Measuring whether the video actually sells

Vanity metrics will mislead you. Autoplay views are not a signal. The numbers that matter sit one step closer to money.

Test one variable at a time. Productive experiments are narrow: first-frame variation, duration variation (six seconds versus twelve), silent versus captioned, product-only versus product-in-use, loop versus single play. Run each against the same traffic split and give it enough time to exit the noise before declaring a winner.

Read the metrics in order:

1. Play rate — did the thumbnail earn the click or the un-mute?
2. Completion to the loop point — did the clip hold attention past the product reveal?
3. Add-to-cart rate, played sessions versus unplayed sessions
4. Return rate — the honest audit of whether the video oversold

If add-to-cart lifts while returns also rise, the video is lying somewhere. The fix is usually in the claim check, not the edit.

There is one structural advantage worth building around: variant production is now cheap. Where a studio test produced one clip, generate four first frames and let the data choose. That is a genuine shift in how creative testing works, and it argues for a pipeline where the master exists once and the variations are all derived from it.

FAQ

Can generated video replace a product photoshoot? No, and it should not try. Real photography anchors accuracy and supplies the reference images generation depends on. Use generation for motion, environments, scale demonstrations, and volume — the shots that would otherwise be too slow or expensive to produce.

How do I stop products from looking subtly wrong? Three habits: a reference image in every job, one motion per shot, and an explicit constraint block in every prompt. Then a shape check against a real photo at the same angle before anything ships. Subtle wrongness is usually a missing constraint, not a weak model.

How long does a first batch take? With assets already collected, a small team can move from shot list to a reviewed batch of eight to twelve clips in about a day: an hour of planning, two to three hours of generation and selection, then assembly. The second batch is faster because the style sheet and templates already exist.

Is this safe for regulated categories? Only with human review at every step. Supplements, cosmetics, medical devices, and financial products carry claim rules a generator cannot know. Keep generation to texture, form, and motion; put every claim in reviewed copy.

Will generated product video hurt search performance? Not inherently. What hurts is a heavy autoplay file that damages load experience, or footage that misrepresents the product and drives returns. Compress the file, keep the claim honest, and make sure the page still has a real video element and a thumbnail if you want it eligible for rich results.

Do I need a different tool per format? Rarely. One pipeline with a fixed style sheet, reusable prompt templates, and clean master exports covers most placements. Switching tools per format is how consistency dies.

What should I generate first if I only have an hour? The mid-motion frame and the texture macro. They are the two shots most catalogs are missing and the two that carry the most information per second.

Put a cinematic pipeline behind your catalog

The pattern that works is unglamorous: one sales question per product, four repeatable formats, a shot list with an honest asset column, prompts written as specifications, a written style sheet, a ninety-second QA pass, and a test that watches add-to-cart instead of views. Do that and generation stops being a novelty and becomes production capacity.

Orelon is built for exactly that kind of work — cinematic ideas in motion, generated from references and precise direction rather than vague wishes. Start with a single hero product and a single format. Generate the shots you cannot photograph, keep the real ones where accuracy matters, and let the AI video generator carry the volume while your judgment carries the standard.