Same product in every shot: how to keep a label from changing in AI video

The product drifts the same way a face does

You render a person holding your jar of hot sauce. Shot one is right. In shot two the lid is a different red. In shot three the label has letters that are almost your brand name and not quite.

That is the same failure as a character's face drifting between scenes, and it has the same fix. An engine that is given no picture of the product invents one from your words, and invents it a little differently every time. Give it the picture and it has something to copy.

The guides on page one of Google agree on the basic advice, and they are right. Livepeer's product-to-video guide says to build a small reference set per product: front, three-quarter and a close-up of the label, shot under the same light. What none of them says is where that pack fits when there is also a person in the shot, or what a wrong label costs you once the render is done. Those two answers come from the reference limits and credit rates we publish, and they decide most of the workflow.

Step 1: shoot the pack once

Take three photos of the real product, or make three clean stills if it does not exist yet:

1. Front, straight on, the whole product in frame.
2. Three-quarter, so the engine learns the depth of the thing: the side of the jar, the thickness of the box.
3. Label close-up, as sharp and as flat as you can get it.

Use plain light and a plain background. A shadow across the label in the reference becomes a smudge on the label in the video. The highest-resolution photo you have is the right one; the engine cannot copy detail it cannot see.

Then write one line describing the product and use it word for word in every prompt, the same way we use a descriptor block for a character. "A squat glass jar, red metal lid, cream label with a green chilli" goes in every prompt, in that order.

Step 2: check how many slots the product takes

This is the step the guides skip. Every engine has a limit on how many references one scene can carry, and the product pack spends some of them.

engine	references per scene	room for a person and a three-view product pack
Seedance 2.5	30 images	yes, with 26 slots left over
Seedance 2.0 and 2.0 Fast	9 images	yes, with 5 left
Kling O3 Pro	5 elements	yes; each element is one subject, so the person is one and the product is another
Veo 3.1 Fast	3 images	no; the pack fills every slot

On Veo 3.1 Fast, three product views leave no room for the person holding it. Either cut the pack to one image (the label close-up is usually the one to keep) or move that shot to another engine. Veo also renders every reference shot at 8 seconds, whatever length you asked for, so a 4-second product close-up there bills 8.

On Kling O3 Pro, a product is an element just like a character. A shot of a host and the product uses two of five elements, which leaves three for a second person, a location or a prop.

Step 3: keep the label big and the shot short

The engines are better at shapes and colours than at small print. A brand name in large letters across the front holds up far better than an ingredients list. Plan your shots around that:

• The label close-up is its own shot, short, with the product filling most of the frame.
• In wide shots the product is a shape and a colour. Nobody reads a label at that size, so do not ask the engine to draw one.
• If the shot needs fine print to be legible, add it in your editor afterwards. invideo's own FAQ calls reference images guidance and warns that small copy can still change.

Step 4: approve a still before you render motion

Make the first frame of each product shot as a still and compare it at full size with the real package. If the label is wrong on the still, it will be wrong in the video, and a still costs far less to redo than a clip.

What a wrong label costs

Each engine on our shelf renders the whole clip at once, so a clip with the wrong label is fixed by rendering it again, billed at the full rate. That makes engine choice a money question:

shot	engine	credits per attempt
4 s close-up, silent	Kling O3 Pro	96
6 s host holds product, voiced	Kling O3 Pro	180
product shot with references, silent	Veo 3.1 Fast (always 8 s)	160
6 s at 480p	Seedance 2.0	132
6 s at 720p	Seedance 2.5	420

A 30-second product ad made of five 6-second shots on Kling O3 Pro, voiced, is 900 credits if every shot lands. If two labels come back wrong, add 360. If you had drafted the label shots at 480p on Seedance 2.0 first, at 132 credits a try, you would know before spending on the final.

This is why we test the label shot first and alone. A product that comes back wrong in a 4-second test costs 96 credits to learn about.

A product in a series

If the same product shows up across a run of videos, a weekly ad or an episode where a character always drinks from the same can, save the product the way you save a character. In our studio you upload the product photos once and save them as an object next to your characters; the prompt then calls it by name, and the next video starts from the same pictures instead of a fresh description. Our guide on AI video for product ads covers the ad formats; this page is only about keeping the thing itself the same.

When the pack is ready, load it into the studio and render the 4-second label close-up before anything else.

FAQ

How do I keep a product the same in every AI video shot?

Give the engine pictures instead of words: front, three-quarter and a label close-up, all in plain light. Use the same one-line product description in every prompt. On ReelDouble you upload the product photos once, save them as an object, and reuse them in every scene.

Why does the label on my product change in AI video?

Without a reference image the engine redraws the product from your text each time, and small print is the first thing it gets wrong. Use a sharp label close-up as a reference, keep label shots short and close, and add fine print in the editor.

Can Veo 3.1 Fast keep a product and a person in the same shot?

Only with one product image. Veo 3.1 Fast takes three reference images, so a three-view product pack fills every slot. On ReelDouble we keep the label close-up and move multi-view product shots to Kling O3 Pro or Seedance.

How much does it cost to fix a wrong label in an AI video?

A wrong label means a new render at the full rate. On ReelDouble a 6-second voiced shot on Kling O3 Pro is 180 credits per try, and a silent 4-second close-up is 96, so test the label shot alone first.

Can I fix just the label without re-rendering the AI video?

Not in the render itself. The practical fix for small text is to add the correct label artwork or text in your editor on top of the finished clip, and to keep fine print out of what the engine has to draw.