Contact us
Product Huntlaunching December 5 Follow the launch

Research · 7 min read

Can AI keep the logo and label on your product? We read every frame

The letters, yes; the label, half the time. We turned photos of four labeled products into 20 videos with five AI video generators and read every frame. Not one letter was misspelled, but 10 of 20 clips broke the label: they added text the product doesn't have, or cut the label off.

A yellow drink can on a dark plinth, its blank label rippling and smearing on one side, framed by a thin green outline of where the label should be

Key takeaways

  • In our November 2026 test, two vision models read 684 lines of label text in 228 frames of 20 AI product videos. Not one line had a misspelled letter.
  • The label still broke in 10 of 20 clips: 6 added fine print, barcodes or a second logo the product doesn't have, and 4 cut the label off, left it blank or moved it to another side.
  • Added text appears on the sides the photo doesn't show. On a drink can, 4 of 5 generators drew fine print or a barcode on its back within 5 seconds.
  • Defects come late: the first one appeared at a median 3.3 seconds into the clip and never in the first second. The first frame of an image-to-video clip is your own photo, so check the last one.
  • The camera move mattered more than the generator: the generator that mostly pushed in kept all 4 labels clean, and the one that turned the camera furthest kept none. That is our reading of 20 clips.

Most AI product videos start from a product photo. An image-to-video generator takes the photo as the first frame and animates it. If the product has a label, a brand manager wants one thing settled before the video goes next to the Add to cart button: does the label still say what it says on the shelf?

We tested it and read every frame.

Can AI video keep the logo and label on a product?

It kept the letters and lost the label. We made four products with brands we invented, each with a three-line label: an olive oil bottle, a serum box, a coffee pouch and a drink can. Five widely used video generators turned each photo into a 5-second clip with the same slow camera move, 20 clips in all. Then two vision models read the label in every frame, every half second.

Not one letter changed. Across 684 readings of a brand name, a product name or a size, no line came back misspelled, garbled or with an invented letter. In 523 of 624 readings after the first frame, the whole line was in view and every letter matched the photo. In the other 101, part of the line had turned out of view.

The label broke anyway. In 10 of 20 clips, something happened to it that a shopper would notice and a brand would reject.

Top row: the first frames of four clips, an olive oil bottle, a serum box twice and a drink can, all with clean labels. Bottom row: the same clips seconds later, with the bottle's label cut to two letters next to a dark band, the box's front panel blank, a second logo on the side of the other box, and fine print and a barcode on the can
Top: the first frame of four clips, which is the product photo itself. Bottom: the same clips seconds later. Generator E cut the bottle's label down to the letters RA next to a dark band and left the box's front panel blank at 4.0 s. Generator C added a second logo on the side of the box. Generator D added a barcode and fine print to the can.

What goes wrong with the label, if not the letters?

Three things went wrong in the 10 broken clips:

  • Text the product doesn’t have (6 clips). As the camera turned, generators drew fine print, barcodes, small icons and, on one box, a second logo on sides the photo never showed.
  • A cut label (3 clips). Two generators shrank the bottle’s label into a strip next to a dark band, so letters vanished while the glass behind them stayed in view. A third turned the coffee pouch sideways and back; it returned with a zipper it never had, and the edge of its own label clipped the brand name.
  • A moved label (1 clip). The text slid off the front of the serum box, the front panel went blank at 4.0 seconds, and the logo reappeared on the side panel.

The drink can shows the pattern most clearly. Its photo has three lines of text on the front and nothing else, because nothing else is in view. Four of five generators filled the back of the can with packaging text they made up.

Six crops of a yellow drink can: the product photo with three lines of text, and the last frame from Generators A to E. A keeps the can as it is; B, C, D and E show columns of fine print on the back of the can, and C and D also show a barcode, E an emblem
The drink can in the photo and at the end of each clip. Generator A pushed in and kept the can as it is. Generators B to E turned it and added fine print, a barcode or an emblem on the back. None of it is on the product.

Our reading: a generator knows only the side of the product in the photo. When the camera turns, it has to draw the rest, and the back of most packaging it has seen carries fine print and a barcode, so that is what it draws. On food and cosmetics, the back of the pack is where the ingredients and legal text go. A video that shows made-up text there shows a product you don’t sell.

Shoppers notice when a product doesn’t match its page. In a survey by Salsify of 1,910 U.S. and U.K. shoppers, 71% said they had returned an item because of incorrect product content.

Which part of the label breaks first: the logo, the name or the small print?

None of them broke by spelling. What differed was how often each line stayed whole in view. After the first frame, the brand name was fully in view and right in 171 of 208 frames, the product name in 157 and the size line in 195.

The product name was the longest line on every label, so it was the first to wrap out of view around a curved bottle or can as the camera turned. The size line was the shortest and sat in the center of the label, so it held up most often. Small type didn’t fail for being small: at 720p, the size line read correctly in every frame where it was in view. We didn’t test anything smaller than a size line, such as an ingredient list.

When does the label start to break?

Late, and that is the trap. The first defect appeared at a median 3.3 seconds into a clip. The earliest came at 1.5 seconds, and none came in the first second. Of the 208 frames after the first, 136 showed the whole label right, 21 had part of a line out of view and 51 showed a defect.

Data table
ProductGeneratorLabelFromWhat happened
Olive oil bottleAClean–Every line right in every frame where it was in view; nothing added.
Olive oil bottleBClean–Every line right in every frame where it was in view; nothing added.
Olive oil bottleCClean–Every line right in every frame where it was in view; nothing added.
Olive oil bottleDLabel cut2.5 sa dark band eats the right part of the label; the first letter disappears while the glass behind it is in view
Olive oil bottleELabel cut2.5 sthe label shrinks into a strip next to a dark band; by the last frame only RA is left
Serum boxAClean–Every line right in every frame where it was in view; nothing added.
Serum boxBClean–Every line right in every frame where it was in view; nothing added.
Serum boxCText added1.5 sa second SOLVEIN logo appears on the side panel
Serum boxDClean–Every line right in every frame where it was in view; nothing added.
Serum boxELabel moved3.5 sthe text slides off the front panel, the front panel is blank at 4.0 s, the logo reappears on the side panel
Coffee pouchAClean–Every line right in every frame where it was in view; nothing added.
Coffee pouchBClean–Every line right in every frame where it was in view; nothing added.
Coffee pouchCClean–Every line right in every frame where it was in view; nothing added.
Coffee pouchDLabel cut4.0 sthe bag turns sideways and back with a new zipper and a flat bottom; the brand word is clipped by its own label edge
Coffee pouchEText added5.9 sfine print appears on the side gusset
Drink canAClean–Every line right in every frame where it was in view; nothing added.
Drink canBText added4.5 scolumns of fine print appear on the side of the can
Drink canCText added3.0 sfine print and a barcode appear on the side
Drink canDText added2.5 sa barcode appears, then fine print and small icons
Drink canEText added3.5 sa block of fine print and an emblem appear on the side
Every square is one frame we read, every half second of every clip. 10 of 20 clips kept the label clean to the last frame; in the other 10, the first defect showed up between 1.5 and 5.9 seconds, never in the first second. Not one of 684 line readings had a misspelled letter. Quelta test, November 6, 2026: five video generators, one clip per product.

So the first frame proves nothing. It is your own photo, so it always passes. A reviewer who checks the opening of a clip and scrubs past the rest would have missed every defect in this test. Shorter clips have less time to drift: in our test, a 2-second clip would have avoided 9 of the 10 defects.

Do some AI video generators keep labels better than others?

In this test, yes, though the camera move explains most of it. Generator A kept all 4 labels clean, B kept 3, C kept 2, D kept 1 and E kept none. Generator A mostly pushed in and barely circled the products, so little of their unseen sides came into view. Generator E turned the camera the furthest: by the end of its clip, the coffee pouch was in profile.

So we don’t read the ranking as “use Generator A.” We read it as “the less of the product a generator has to invent, the more of the label survives.” The products point the same way. The can, round and with the most new surface turning into view, stayed clean in 1 of 5 clips; the bottle, the box and the pouch in 3 of 5 each. That part is our reading of the results. The ranking itself is a snapshot of November 2026, one clip per product and generator.

How to check the label in an AI product video before it goes live

Read it frame by frame, at full size. These six checks take a couple of minutes per clip:

Tick each check the video passes.
Six label checks for every AI product video, from our test. Tick the ones your video passes.

If you make the videos yourself, two choices help. For shots where the label has to be read, use a push-in or a short pan rather than an orbit. And if the camera has to show the back or the sides, remember that without a photo of them, the generator will draw them.

The file itself, its length, format and resolution for each marketplace, is a separate check: our free video requirements checker runs it in the browser.

How we keep the label right in production

The same way we keep the size right, which we covered in why AI product videos get the size wrong: we check every frame. Our engine builds each video from the product’s photos and checks every frame against them for size, shape, the number of parts, the logo and text, and color. Text the product doesn’t have fails the check, the same as a misspelled logo. A frame that fails is regenerated before the video is assembled.

STONE HEART, an isotonic drink brand we made up for our demos, in an ad made with our engine: the same label on the bottle in every shot it appears in, through to the final packshot.

More videos made this way are on our samples page.

What to ask before you order AI product video

Four questions show whether a vendor reads the label or hopes for it.

Question A good answer A red flag
How do you check the logo and text? In every frame, against the product photos “Our team watches every video”
What appears on the sides my photos don’t show? Only what is on the real product “The generator fills them in”
What happens to a frame with text the product doesn’t have? It is regenerated before you see the video You find it in review and ask for a revision
Can you show the last frame of each shot at full size? Yes, for every shot Only the video, at phone size

If you want to see the answers on your own catalog, we start with a pilot on 3⁠–⁠5 products: send a link to your catalog, and read the labels in our videos against your own packaging. How this works for a whole catalog is on our page for marketplaces.

Sources

  1. Quelta label text test, November 6, 2026: every line read in every frame (CSV)
  2. Salsify, 2025 Consumer Research Executive Summary (survey of 1,910 U.S. and U.K. shoppers, October 2024)

Quelta for catalogs

Need videos like this for your whole catalog?

Quelta makes product videos from your photos, adapted for each of your markets by meaning, not just by words, and checked before delivery. Start with a pilot on one category.

Start with a pilot

Tell us about your catalog

Send a link to your catalog or website and tell us roughly how many videos you need a month. We start with a pilot on 3⁠–⁠5 of your products, so you can judge the videos on your own catalog.

  1. Send a link to your catalog

    Your website, a product feed or your store on a marketplace.

  2. We make a pilot on 3⁠–⁠5 of your products

    Product-accurate videos of your real products, for the markets and platforms you need.

  3. A monthly contract for the volume you need

    New videos for the catalog and new arrivals every month, in one style.

We don’t copy other people’s videos. Every video is new and about your product. We keep the mechanics that sell (script, pacing, delivery), not anyone’s footage, people or music.

Contact us

Prefer to talk first? Yegor, Co-founder. A 30-minute video call about your catalog.

Book a call

Made with our engineTests, not client projects

Questions and answers

Can AI video keep the logo on my product correct?

It keeps the letters, not always the label. In our November 2026 test of 20 image-to-video clips, no line of label text was misspelled in 684 readings. But 10 of 20 clips broke the label in other ways: they added text the product doesn't have, cut the label off or moved the logo to another side.

Why does AI video add text that isn't on my product?

A generator only sees the side of the product in your photo. When the camera turns, it draws the rest from what packaging usually looks like, which often means fine print and a barcode. In our test, 4 of 5 generators added such text to the back of a drink can within 5 seconds.

Does image-to-video keep the text from my product photo?

Yes, while that text stays in view. The first frame is your photo, and in our test the generators copied its letters without a single misspelling. The risk starts when the camera shows a side the photo didn't, or turns far enough for the label to wrap out of view.

How long can an AI product video run before the label breaks?

In our test, the first label defect appeared at a median 3.3 seconds into a clip, the earliest at 1.5 seconds. A 2-second clip would have avoided 9 of the 10 defects we found. Anything longer needs a frame-by-frame check.

How do I check the label in an AI product video?

Pause on the last frame and on the frame where the camera has turned the most, and read every line at full size. Then look at the sides your photo didn't show for fine print, barcodes or a second logo, and check that the label's edges wrap around the product.