Consistent AI product video: keep the product the product
A product in an AI video is consistent when four things hold in every shot: the shape, the label, the logo and the color. Each one fails for a different reason, so each needs a different fix. This page breaks product consistency into those four parts, explains what current reference models get right and wrong as of October 2026, and gives a shot plan for ads and demos where an inaccurate product is not an option.
Why products are harder than faces
Viewers forgive a face that is a little off. They do not forgive a bottle that changes shape between two cuts, or a logo with an extra letter. Products are also often reflective, transparent or covered in small print, which are the things video models handle worst.
There is a business problem too. An ad that shows the product inaccurately can mislead buyers. Treat accuracy as a requirement, not a bonus.
Shape: proportions and silhouette
Shape drifts when the model sees the product from one angle and has to guess depth. A front-only reference turns a deep box into a flat one, or a round bottle into an oval one, the moment the camera orbits.
Fix: give at least three angles, such as front, side and a three-quarter view from above, plus one image with a hand for scale. Avoid wide-angle phone shots taken up close, because lens distortion becomes part of the shape. In the prompt, describe size relative to the scene, such as "fits in one hand", rather than the shape itself.
Label: text and layout
Every current video model redraws small text. Large brand names often hold. Ingredient lists, sizes and fine print usually come back as plausible scribbles. References cannot fully solve this today.
Fix: decide which text must be legible and plan around it. Use one hero close-up where the main label faces the camera squarely and the camera moves slowly. In other shots, keep the product at mid-distance or turned slightly, so the label is visible but not readable. If exact wording must appear, add it as a graphic in editing or cut to a real packshot.
Logo: marks and placement
Logos fail in two ways. The mark itself changes, with letters merging or a symbol simplifying. Or it moves on the product, for example from the cap to the body. The first is a resolution problem. The second is a coverage problem.
Fix: include one sharp close-up where the logo fills a good part of the frame, and make sure the wider references show where it sits. Use the highest resolution your tool offers for logo shots, and keep motion slow while the logo is in view. Fast pans and motion blur are where marks break.
Color: hue and finish
Color drifts because the scene's light is applied to the product. A coral bottle in a sunset scene comes back orange. A matte black device under neon light turns glossy blue. The model is not wrong about physics, but the product no longer matches the shelf.
Fix: shoot references in neutral, even light with no filters and correct white balance, so the reference color is the real color. Use neutral or soft light for product-focused shots and save dramatic color for shots where the product is small in frame. For brand-critical colors, plan a light grade in editing and compare with a photo of the real product.
A shot plan for a 15-second product ad
A layout that keeps risk low: a 3-second shot of the product in use at mid-distance, label angled away. A 5-second slow push-in to the hero angle with the main label square to the camera. A 3-second detail of texture, cap or button. Then a real packshot or end card with exact text, made in your editor.
Generate every shot from the same product references and the same model. Hands or an unnamed person can be described in the prompt. If a recurring presenter must hold the product, you need a model that takes two subjects in one request, such as Kling's Omni models with two elements.
Before you publish
Pause every shot where the product is visible and check the four points against the physical product: silhouette, label layout, logo and color. Check reflections and transparent parts for impossible geometry. Remove any shot where text is readable but wrong.
On RefVid, a product is an asset with 1 to 5 reference images, the main view first. Every shot of a campaign starts from that same set on Kling 3.0 Omni, and all clips sit on the asset's page for this review. You can rate each clip, so the takes worth keeping are easy to find.
Frequently asked questions
- Can AI video keep a product label readable?
- Large brand names often hold. Small text such as ingredients and sizes is usually redrawn. Plan one slow hero close-up for the main label and add exact wording in editing.
- Why does my product change color in some scenes?
- The scene's lighting is applied to it. Shoot references in neutral light, use soft light for product-focused shots, and check color against the real product before publishing.
- How many product reference images do I need?
- Three to five: front, side, a three-quarter view from above, a logo close-up, and one image that shows scale. Add the back if it ever appears on screen.
- What does a product shot cost on RefVid?
- 42 credits per second: 126 for a 3-second shot, 210 for 5 seconds and 420 for 10. The first 50 accounts get 630 free credits. Accounts after that get 126.