Lock the product still, then run I2V

The still is the contract. Approve the pack shot — label, colourway, angle — then run image-to-video from that frame. Do not text-to-video a SKU you already photographed. T2V will invent a cousin bottle. I2V will move the bottle you signed off.

This is the product path: still, then I2V. It is not motion control (that needs a driving clip). It is not a 30-second storyboard.

The still is the SKU contract

A product ad has one object that cannot drift: the thing for sale. Legal, marketplace, and the buyer all need that object. A generate that almost looks like the SKU is a miss even when the motion is good.

Lock the still until someone can freeze the frame and say “that is the listing.”

Checklist for the still:

The actual SKU. Not a moodboard cousin. Not last year’s colourway. The barcode-level product if the campaign is that tight.
Label lettering you can read. If the still already mush-letters the flavour name, I2V will mush it in motion. Fix the still with an image model or a photograph. Do not hope video will typeset.
Angle that matches the verb. A top-down still is a bad contract for a three-quarter hero turn. A beauty crop with no cap is a bad contract for an unscrew. Shoot or generate the still in the pose the video must start from.
Ratio you will ship. Crop the still to 9:16 before I2V if the clip is for TikTok / Reels / Shorts. A 16:9 pack on a table becomes a floating label when you later crop. I2V follows the still.
Background you can live with. I2V will invent motion in the room. A messy still becomes a messy room that moves.
Rights. You own the photo or the generate. A scraped PDP image with a watermark is not a contract.

Iterate at image prices until the freeze-frame passes. That is cheaper than a 70-credit I2V on the wrong bottle.

What “locked” means in practice: the file is named, in the brief, and nobody is still “trying a different angle in the prompt.” The prompt on I2V is the motion, not a second chance at identity.

If you do not have a still, you do not have this job. You have a stills job. Do that first. T2V is allowed when there is no SKU — a mood, a city, a generic cup of coffee. The moment a label exists, stop T2V.

Pick the I2V row for the motion you need

Attach the locked still. Then pick a published image-to-video (or reference) row from the motion and audio you need.

Examples on this catalog (credits are the card, not USD):

Kling O3 Pro Image to Video — I2V, native audio, 3–15s, 70 credits. When the still must speak and you want that family’s motion.
Wan 2.7 Image to Video — I2V, 15 credits, 1080p, 2–15s, audio null on this card. Do not claim native audio. Silent plate, then TTS if you need VO.
Seedance 2.0 — T2V + R2V, native audio, 1080p, 4–15s, 69 credits. When you have references (the still plus others) rather than a single I2V frame.
Seedance 2.0 Fast Reference to Video — R2V, native audio, 720p, 4–15s, 25 credits. Cheaper reference path when 720p is enough.
Veo first-last-frame — I2V-shaped, native audio, 4K, 4/6/8s, 40 credits. When you can lock start and end stills.

The AI video generator will not switch you from T2V to I2V because the prompt mentioned a brand name. You attach the frame.

Procedure:

Open the I2V / R2V row, not the T2V twin.
Attach the locked still (and extra references on R2V).
Prompt only the motion: slow orbit, hands enter from the right, cap unscrews, liquid pours, camera holds on the label. Do not re-describe the bottle (“sleek teal serum with gold cap”) as if the still were missing. The still is the bottle.
Set duration to one verb. A 15-second prompt with five verbs is how the SKU teleports. One generate, one beat.
Set aspect to the still’s aspect.
Generate. Freeze several frames. Compare to the still. If the label drifted, the take failed — even if the motion is pretty.
If audio is native and you did not want speech, say so in the prompt (silent room, no voiceover) or pick a silent I2V card. Baked speech on a product orbit is extra work to replace.

Do not pick T2V “just to see.” You will see a cousin. You will then try to match it in Photoshop. That is how a week disappears.

Do not T2V a SKU

T2V is a sentence about a product the model has never held. It will get the silhouette in the neighbourhood and the label in another language. That is expected. It is not a bug you fix with more adjectives.

Forbidden shortcuts:

“Exact product, do not change the label” in a T2V prompt. The model cannot see the label you did not attach.
Screenshot of a competitor as “reference” without a row that actually takes images. T2V will ignore the screenshot if the card has no image slot.
I2V from a still you have not locked, then blaming the video model for the wrong colourway. The still was the miss.
Motion-control with a product still and a dance driving clip when you wanted an orbit. Motion control copies the drive. You will get a dancing bottle. Use I2V for product motion you can describe, or a turntable driving clip if you truly want that path.

When T2V is allowed next to a SKU campaign: B-roll that is not the pack (steam, city, hands without the label), or a hook face that never holds the product. The moment the pack is on screen, return to the still.

UGC that shows a product follows the same rule. A talking head can be T2V or lipsync. The insert of the bottle is I2V from the locked still. Two rows, one brief.

QC the I2V against the still, then caption

Score the take as a listing, not as cinema.

Pause on three frames: start, middle, end. Side-by-side with the still.
Read the label out loud from the video frames. If you cannot, the still was too soft or the motion too violent.
Check colourway under the same grade. A warmer I2V is a different SKU on some shelves.
Check geometry: cap, pump, carton edges. Morphing corners are a failed contract.
Check extra logos and extra packs the model invented in the background. Crop or regenerate.
Only then add native audio replacement or TTS if the card was silent, and burn captions. Do not caption a cousin bottle.

If middle-frame drift is the failure, shorten the duration and reduce camera travel. A 4-second hold with a tiny push-in keeps more identity than a 15-second orbit.

If you need a studio seat to attach the still and run I2V, sign up. Bring the locked frame. Do not bring a paragraph about the bottle.

FAQ

Why not describe the product in T2V if I do not have a photo yet?

Then you still do not have a SKU contract. Generate or shoot the still first, approve the label, then I2V. T2V is how you get a cousin you cannot list.

Which I2V row keeps the pack most faithfully?

Identity comes from the still and from not asking five verbs in one generate. Pick Kling O3 Pro I2V, Wan 2.7 I2V, or a Seedance reference row from audio, duration, and resolution — then QC against the frame.

Can I T2V the hook and I2V the pack in the same ad?

Yes. That is a storyboard: face shot on one row, pack insert from the locked still on I2V, concatenate, caption once. Do not ask one T2V pass to do both.